Tin-Yau Tam Multilinear Algebra (Lecture Notes)

Multilinear Algebra1
Tin-Yau Tam
Department of Mathematics and Statistics
221 Parker Hall
Auburn University
AL 36849, USA
tamtiny@auburn.edu
November 30, 2011
1
Some portions are from B.Y. Wangs Foundation of Multilinear Algebra (1985 in
Chinese)
2
Chapter 1
Review of Linear Algebra
1.1 Linear extension

In this course, U, V, W are finite dimensional vector spaces over C, unless spec-
ified. All bases are ordered bases.
Denote by Hom (V, W ) the set of all linear maps from V to W and End V :=
Hom (V, V ) the set of all linear operators on V . Notice that Hom (V, W ) is a vec-
tor space under usual addition and scalar multiplication. For T Hom (V, W ),
the image and kernel are
Im T = {T v : v V } W, Ker T = {v : T v = 0} V
which are subspaces. It is known that T is injective if and only if Ker T = 0.

The rank of T is rank T := dim Im T .
Denote by Cmn the space of m n complex matrices. Each A Cmn can
be viewed as in Hom (Cn , Cm ) in the obvious way. So we have the concepts, like
rank, inverse, image, kernel, etc for matrices.
Theorem 1.1.1. Let E = {e1 , . . . , en } be a basis of V and let w1 , . . . , wn W .

Then there exists a unique T Hom (V, W ) such that T (ei ) = wi , i = 1, . . . , n.
Pn Pn Pn
Proof. For each v = i=1 ai ei , define T v = i=1 ai T ei = i=1 ai wi . Such
T is clearlyP
linear. If S, TP Hom (V, W ) such that T (eP i ) = wi , i = 1, . . . , n,
n n n
then Sv = i=1 ai Sei = i=1 ai T ei = T v for all v = i=1 ai ei V so that
S = T.
In other words, a linear map is completely determined by the images of basis

elements in V .
A bijective T Hom (V, W ) is said to be invertible and its inverse (T 1
T = IV and T T 1 = IW ) is linear, i.e., T 1 Hom (W, V ): T 1 (1 w1 +
2 w2 ) = v means that
T v = 1 w1 + 2 w2 = T (1 T 1 w1 + 2 T 1 w2 ),
3
4 CHAPTER 1. REVIEW OF LINEAR ALGEBRA
i.e., v = T 1 w1 + 2 T 1 w2 for all w1 , w2 W and v V . We will simply write

ST for S T for S Hom (W, U ), T Hom (V, W ). Two vector spaces V and
W are said to be isomorphic if there is an invertible T Hom (V, W ).
Theorem 1.1.2. Let T Hom (V, W ) and dim V = n.
1. Then rank T = k if and only if there is a basis {v1 , . . . , vk , vk+1 , . . . , vn }
for V such that T v1 , . . . , T vk are linearly independent and T vk+1 = =
T vn = 0.
2. dim V = dim Im T + dim Ker T.
Proof. Since rank T = k, there is a basis {T v1 , . . . , T vk } for Im T . Let {vk+1 , . . . , vk+l }
be a basis of Ker T . Set E = {v1 , . . . , vk , vk+1 , . . . , vk+l }. For each v V ,
Pk Pk Pk
T v = i=1 ai T vi since T v Im T . So T (v i=1 ai vi ) = 0, i.e., v i=1 ai vi
Pk Pk+l Pk+l
Ker T . Thus v i=1 ai vi = i=k+1 ai vi , i.e., v = i=1 ai vi so that E spans
Pk+l
V . So it suffices to show that E is linearly independent. Suppose i=1 ai vi = 0.
Pk
Applying T on both sides, i=1 ai T vi = 0 so that a1 = = ak = 0. Hence
Pk+l
i=k+1 ai vi = 0 and we have ak+1 = = ak+l = 0 since vk+1 , . . . , vk+l are
linear independent. Thus E is linearly independent and hence a basis of V ; so
k + l = n.
Theorem 1.1.3. Let A Cmn .
1. rank A A = rank A.
2. For each A Cmn , rank A = rank A = rank AT , i.e., column rank and
row rank of A are the same.
Proof. (1) Notice Ax = 0 if and only if A Ax = 0. It is because A Ax = 0
implies that (Ax) (Ax) = x A Ax = 0. So Ax1 , . . . , Axk are linearly in-
dependent if and only if A Ax1 , . . . , A Axk are linearly independent. Hence
rank A A = rank A.
(2) Since rank A = rank A A rank A from Problem 3 and thus rank A
rank A since (A ) = A. Hence rank A = rank AT (why?).
Problems
1. Show that dim Hom (V, W ) = dim V dim W .
2. Let T Hom (V, W ). Prove that rank T min{dim V, dim W }.
3. Show that if T Hom (V, U ), S Hom (U, W ), then rank ST min{rank S, rank T }.
4. Show that the inverse of T Hom (V, W ) is unique, if exists. Moreover T
is invertible if and only if rank T = dim V = dim W .
5. Show that V and W are isomorphic if and only dim V = dim W .
1.2. MATRIX REPRESENTATIONS OF LINEAR MAPS 5
6. Show that if T Hom (V, U ), S Hom (U, W ) are invertible, then ST is

also invertible. In this case (ST )1 = T 1 S 1 .
7. Show that A Cmn is invertible only if m = n, i.e., A has to be square.

Show that A Cnn is invertible if and only if the columns of A are
linearly independent.
8. Prove that if A, B Cnn such that AB = In , then BA = In .
Solutions to Problems 1.1
1. Let {e1 , . . . , en } and {f1 , . . . , fm } be bases of V and W respectively. Then

ij Hom (V, W ) defined by ij (ek ) = ik fj , i, j, k = 1, . . . , n, form a basis
of Hom (V, W ). Thus dim Hom (V, W ) = dim V dim W .
2. From definition rank T dim W . From Theorem 1.1.2 rank T dim V .
3. Since Im ST Im S, rank ST rank S. By Theorem 1.1.2 rank ST =

dim V dim Ker ST . But Ker T Ker ST so that rank ST dim V
dim Ker T = rank T .
4. Suppose that S and S 0 Hom (V, W ) are inverses of T Hom (V, W ).

Then S = S(T S 0 ) = (ST )S 0 = S 0 ; inverse is unique. T Hom (V, W )
invertible T is bijective (check!). Now injective T Ker T = 0, and
surjective T rank T = dim W
5. One implication follows from Problem 4. If dim V = dim W , let E =

{e1 , . . . , en } and {f1 , . . . , fn } be bases of V and W , define T Hom (V, W )
by T ei = fi for all i which is clearly invertible since T 1 fi = ei .
6. (T 1 S 1 )(ST ) = IV and (ST )(T 1 S 1 ) = IW .
7. From Problem 4, a matrix A is invertible only if A is square. The second

statement follows from the fact the rank A is the dimension of the column
space of A.
8. If AB = I, then rank AB = n so rank A = n(= rank B) by Problem

3. So Ker A = 0 and Ker B = 0 by Theorem 1.1.2. Thus A and B is
invertible. Then I = A1 ABB 1 = A1 B 1 (BA)1 so that BA = I.
(or, without using Theorem 1.1.2, note that B = B(AB) = (BA)B so
that (I BA)B = 0. Since Im B = Cn , I BA = 0, i.e., BA = I).
1.2 Matrix representations of linear maps

In this section, U, V, W are finite dimensional vector spaces.
Matrices A Cmn can be viewed as elements in Hom (Cn , Cm ). On the
other hand, each T Hom (V, W ) can be realized as a matrix once we fix bases
for V and W .
Let T Hom (V, W ) with bases E = {e1 , . . . , en } for V and F = {f1 , . . . , fm }

for W . Since T ej W , we have
m
X
T ej = aij fi , j = 1, . . . , n.
i=1
The matrix A = (aij ) Cmn is called the matrix representation of T with

respect to the bases E and F , denoted by [T ]F E = A.
From Theorem 1.1.1 [S]F F
E = [T ]E if and only if S = T , where S, T
Hom (V, W ).
The coordinate vector of v V with respect to the basis Pn E = {e1 , . . . , en }
for V is denoted by [v]E := (a1 , . . . , an )T Cn where v = i=1
Pani vi . Indeed we
can view v Hom (C, V ) with 1 as a basis of C. Then v 1 = i=1 ai ei so that
[v]E is indeed [v]E
1.
Theorem 1.2.1. Let T Hom (V, W ) and S Hom (W, U ) with bases E =
{e1 , . . . , en } for V , F = {f1 , . . . , fm } for W and G = {g1 , . . . , gl } for U . Then
[ST ]G G F
E = [S]F [T ]E
and in particular
[T v]F = [T ]F
E [v]E ,
v V.
Pm
Proof. Let A = [T ]F G
E and B = [S]F , i.e., T ej = i=1 aij fi , j = 1, . . . , n. and
Pl
T fi = k=1 bki gk , i = 1, . . . , m. So
m
X m
X l
X l
X
ST ej = aij Sfi = aij ( bki )gk = (BA)kj gk .
i=1 i=1 k=1 k=1
So [ST ]G G F
E = BA = [S]F [T ]E .
0
Consider I := IV End V . The matrix [I]E
E is called the transitive matrix
from E to E 0 since
0
[v]E 0 = [Iv]E 0 = [I]E
E [v]E , v V,
0
i.e., [I]E
E transforms the coordinate vector [v]E with respect to E to [v]E 0 with
E 0 1
respect to E 0 . From Theorem 1.2.1, [I]E
E 0 = ([I]E ) .
Two operators S, T End V are said to be similar if there is an invertible
P End V such that S = P 1 AP . Similarity is an equivalence relation and is
denoted by .
Theorem 1.2.2. Let T End V with dim V = n and let E and E 0 be bases
0
of V . Then A := [T ]E E
E 0 and B := [T ]E are similar, i.e., there is an invertible
1
P Cnn such that P AP = B. Conversely, similar matrices are matrix
representations of the same operator with respect to different bases.
1.2. MATRIX REPRESENTATIONS OF LINEAR MAPS 7
0
Proof. Let I = IV . By Theorem 1.2.1, [I]E E E
E 0 [I]E = [I]E = In where In is n n
identity matrix. Denote by P := [I]E 0 (the transitive matrix from E 0 to E) so
E
0
that P 1 = [I]E
E . Thus
0 0 0
[T ]E E E E E
E 0 = [IT I]E 0 = [I]E [T ]E [I]E 0 = P
1
[T ]E
E P.
Suppose that A and B are similar, i.e., B = R1 BR. Let E be a basis of V .

By Theorem 1.1.1 A uniquely determines T End V such that [T ]E
E = A. By
Theorem 1.2.1, an invertible R Cnn uniquely determines a basis E 0 of V
such that [T ]E
E 0 = R so that
0 0
[T ]E E E E
E 0 = [I]E [T ]E [I]E 0 = R
1
[T ]E
E R = B.
So functions : Cnn C that take constant values on similarity orbits

of Cnn , i.e., (A) = P 1 AP for all invertible P Cnn , are defined for
operators, for examples, determinant and trace of a operator.
Problems
1. Let E = {e1 , . . . , en } and F = {f1 , . . . , fm } be bases for V and W . Show
that the matrix representation := [ ]F E : Hom (V, W ) Cmn is an
isomorphism.
2. Let E and F be bases of V and W respectively. Show that if T
Hom (V, W ) is invertible, then [T 1 ]E F 1
F = ([T ]E ) .
3. Let E and F be bases of V and W respectively. Let T Hom (V, W ) and
A = [T ]F
E . Show that rank A = rank T .

1. It is straightforward to show that is linear by comparing the (ij) entry of
[S +T ]F F F
E and [S]E +[T ]E . Since dim Hom (V, W ) = mn = dim Cmn ,
it suffices to show that is injective. From Theorem 1.1.1, is injective.
2. From Theorem 1.2.1, [T 1 ]E F
F [T ]E = [T
1
T ]E E
E = [IV ]E = In where dim V =
n. Thus use Problem 1.8 to have [T ]F = ([T ]F
1 E
E)
1
, or show that
F 1 E
[T ]E [T ]F = In similarly.
3. (Roy) Let rank T = k and let E = {v1 , . . . , vn } and F = {w1 , . . . , wm } be
bases for V and W . When we view v V as an element in Hom (C, V ),
by Problem 1, the coordinate maps [ ]E : V Cn and [ ]F : W Cm
are isomorphisms. So
rank T = dim Im T = dimhT v1 , . . . , T vn i = dimh[T v1 ]F , . . . , [T vn ]F i
= dimh[T ]F F
E [v1 ]E , . . . , [T ]E [vn ]E i = dim Im A = rank A.
1.3 Inner product spaces

Let V be a vector space. An inner product on V is a function (, ) : V V C
such that
1. (u, v) = (v, u) for all u, v V .
2. (1 v1 + 2 v2 , u) = 1 (v1 , u) + 2 (v2 , u) for all v1 , v2 , u V , 1 , 2 C.
3. (v, v) 0 for all v V and (v, v) = 0 if and only if v = 0.
The space V is then called an inner product space. The norm induced by
the inner product is defined as
p
kvk = (v, v), v V.
Vectors v satisfying kvk = 1 are called unit vectors. Two vectors u, v V are
said to be orthogonal if (u, v) = 0, denoted by u v. A basis E = {e1 , . . . , en }
is called an orthogonal basis if the vectors are orthogonal. It is said to be
orthonormal if
(ei , ej ) = ij , i, j = 1, . . . , n.
where (
1 if i = j
ij =
0 if i =
6 j
is the Kronecker delta notation.
Theorem 1.3.1. (Cauchy-Schwarz inequality) Let V be an inner product space.
Then
|(u, v)| kukkvk, u, v V.
Equality holds if and only if u and v are linearly dependent, i.e., one is a scalar
multiple of the other.
(u,v)
Proof. It is trivial when v = 0. Suppose v 6= 0. Let w = u kvk2 v. Clearly
(w, v) = 0 so that
(u, v) |(u, v)|2
0 (w, w) = (w, u) = (u, u) 2
(v, u) = kuk2 .
kvk kvk2
Equality holds if and only if w = 0, i.e. u and v are linearly dependent.
Theorem 1.3.2. (Triangle inequality) Let V be an inner product space. Then
ku + vk kuk + kvk, u, v V.
Proof.
ku + vk2 = (u + v, u + v) = kuk2 + 2Re (u, v) + kvk2 kuk2 + 2|(u, v)| + kvk2
By Theorem 1.3.1, we have
ku + vk2 kuk2 + 2kukkvk + kvk2 = (kuk + kvk)2 .
1.3. INNER PRODUCT SPACES 9
Theorem 1.3.3. Let E = {e1 , . . . , en } be an orthonormal basis of V . For any

u, v V ,
n
X
u = (u, ei )ei ,
i=1
Xn
(u, v) = (u, ei )(ei , v).
i=1
Pn Pn
Proof. Let u = j=1 aj ej . Then (u, ei ) = ( j=1 aj ej , ei ) = ai , i = 1, . . . , n.
Pn Pn
Now (u, v) = ( i=1 (u, ei )ei , v) = i=1 (u, ei )(ei , v).
Denote by hv1 , . . . , vk i the span of the vectors v1 , . . . , vk .
Theorem 1.3.4. (Gram-Schmidt orthogonalization) Let V be an inner product

space with basis {v1 , . . . , vn }. Then there is an orthonormal basis {e1 , . . . , en }
such that
hv1 , . . . , vk i = he1 , . . . , ek i, k = 1, . . . , n.
Proof. Let
v1
e1 = ,
kv1 k
v2 (v2 , e1 )e1
e2 =
kv2 (v2 , e1 )e1 k
...
vn (vn , e1 )e1 (vn , en1 )en1
en =
kvn (vn , e1 )e1 (vn , en1 )en1 k
It is direct computation to check that {e1 , . . . , en } is the desired orthonormal

basis.
The inner product on Cn : for any x = (x1 , . . . , xn )T and y = (y1 , . . . , yn )T

n
C ,
n
X
(x, y) := xi yi
i=1
is called the standard inner product on Cn . The induced norm is called the
2-norm and is denoted by
kxk2 := (x, x)1/2 .
Problems
1. Let E = {e1 , . . . , en } be a basis of V and v V . Prove that v = 0 if and

only if (v, ei ) = 0 for all i = 1, . . . , n.
2. Show that each orthogonal set of nonzero vectors is linearly independent.

Pn
3. Let E
P= {e1 , . . . , en } be an orthonormal
P basis of V . If u = i=1 ai ei and
n n
n = i=1 bi ei , then (u, v) = i=1 ai bi .
4. Show that an inner product is completely determined by an orthonormal

basis, i.e., if the inner products (, ) and h, i have the same orthonormal
basis, then (, ) and h, i are the same.
5. Show that (A, B) = tr (B A) defines an inner product on Cmn , where

B denotes the complex conjugate transpose of B.
6. Prove |tr (A B)| tr (A A)tr (B B) for all A, B Cmn .
1. v = 0 (v, u) = 0 for allPu V (v, ei ) = 0 for all i. Conversely

n
each u V is of the form i=1 ai ei so that (v, ei ) = 0 for all i implies
(v, u) = 0 for all u V .
2. Suppose
Pnthat S = {v1 , . . . , vn } isP an orthogonal set of nonzero vectors.
n
With i=1 ai vi = 0, aj (vj , vj ) = ( i=1 ai vi , vj ) = 0 so that aj = 0 for all
j since (vj , vj ) 6= 0. Thus S is linearly independent.
3. Follows from Theorem 1.3.3.
4. Follows from Problem 3.
5. Straightforward computation.
6. Apply Cauchy-Schwarz inequality on the inner product defined in Problem

5.
1.4 Adjoints
Let V, W be inner product spaces. For each T Hom (V, W ), the adjoint of T
is S Hom (W, V ) such that (T v, w)W = (v, Sw)V for all v V, w W and is
denoted by T . Clearly (T ) = T .
Theorem 1.4.1. Let W, V be inner product spaces. Each T Hom (V, W ) has
a unique adjoint.
Proof. By Theorem 1.3.4, let E = {e1 , . . . , en } be an orthonormal basis of V .

For any w W , define S Hom (W, V ) by
n
X
Sw := (w, T ei )W ei .
i=1
1.4. ADJOINTS 11
Pn
For any v V , by Theorem 1.3.3, v = i=1 (v, ei )V ei so that
n
X n
X
(v, Sw)V = (v, (w, T ei )W ei )V = (T ei , w)W (v, ei )V
i=1 i=1
Xn
= (T (v, ei )V ei , w)W = (T v, w)W .
i=1
Uniqueness follows from Problem 1.
Theorem 1.4.2. Let E = {e1 , . . . , en } and F = {f1 , . . . , fm } be orthonormal

bases of the inner product spaces V and W respectively. Let T Hom (V, W ).
Then [T ]E F
F = ([T ]E ) , where the second denotes the complex conjugate trans-
pose.
Pm
Proof. Let A := [T ]F E , i.e., T ej = k=1 akj fk , j = 1, . . . , n. So (T ej , fi )W =
aij . By the proof of Theorem 1.4.1
n
X n
X n
X
T fj = (fj , T ei )W ei = (T ei , fj )W ei = aji ei .
i=1 i=1 i=1
So [T ]E F
F = A = ([T ]E ) .
Notice that if E and F are not orthonormal, then [T ]E F

F = ([T ]E ) may not
hold.
Problems
1. Let S, T Hom (V, W ) where V, W are inner product spaces. Prove that
(a) (T v, w) = 0 for all v V and w W if and only if T = 0.

(b) (Sv, w) = (T v, w) for all v V and w W if and only if S = T .
2. Show that (ST ) = T S where T Hom (V, W ) and S L(W, U ) and

U, V, W are inner product spaces.
3. Let E and F be orthonormal bases of inner product space V . Prove that

([I]F F 1
E ) = ([I]E ) = [I]E
F.
4. Let G be a basis of the inner product space V . Let T End V . Prove

that [T ]G G
G and ([T ]G ) are similar.
5. Let V, W be inner product spaces. Prove that if T Hom (V, W ), then

rank T = rank T .
6. The adjoint : Hom (V, W ) Hom (W, V ) is an isomorphism.

1. (a) (T v, w) = 0 for all v V and w W T v = 0 for all v V , i.e.,

T = 0.
(b) Consider S T .
2. Since (v, T S w)V = (T v, S w)V = (ST v, w)V , by the uniqueness of ad-

joint, (ST ) = T S .
3. Notice that I = I where I := IV . So from Theorem 1.4.2 ([I]F

E) =
E F F 1 E
[I ]F = [I]E . Then by Problem 2.2, ([I]E ) = [I]F .
4. (Roy and Alex) Let E be an orthonormal basis of V . Let P = [IV ]E G

(the transition matrix from G to E). Then [T ]G G = P
1
[T ]E
E P and
[T ]E G 1
E = P [T ]G P . By Theorem 1.4.2 [T ]E E
E = ([T ]E ) . Together with
Problem 2
[T ]G
G = P 1 [T ]E
EP = P
1
([T ]E
E) P
= P 1 (P [T ]GGP
1
) P = (P P )1 ([T ]G
G ) (P P ),
i.e. ([T ]G G
G ) and ([T ]G ) are similar.
Remark: If G is an orthonormal basis, then P is unitary and thus ([T ]G
G) =
G
([T ]G ), a special case of Theorem 1.4.2.
5. Follows from Theorem 1.4.2, Problem 2.3 and rank A = rank A where A
is a matrix.
6. It is straightforward to show that : Hom (V, W ) Hom (W, V ) is a

linear map. Since dim(V, W ) = dim(W, V ) it remains to show that is
injective. First show that (T ) = T : for all v V , w W ,
(T w, v) = (v, T w)V = (T v, w)V = (w, T v)
Then for any S, T Hom (V, W ), S = T S = T .
1.5 Normal operators and matrices

An operator T End V on the inner product space V is normal if T T = T T ;
Hermitian if T = T , positive semi-definite, abbreviated as psd or T 0
if (T x, x) 0 for all x V ; positive definite, abbreviated as pd or T > 0 if
(T x, x) > 0 for all 0 6= x V ; unitary if T T = I. Unitary operators form a
group.
When V = Cn is equipped with the standard inner product and orthonormal
basis, linear operators are viewed as matrices in Cnn and the adjoint is simply
the complex conjugate transpose. Thus a matrix A Cnn is said to be normal
if A A = AA ; Hermitian if A = A , psd if (Ax, x) 0 for all x Cn ; pd
if (Ax, x) > 0 for all 0 6= x Cn ; unitary if A A = I. Unitary matrices in
1.5. NORMAL OPERATORS AND MATRICES 13
Cnn form a group and is denoted by Un (C). One can see immediately that
A is unitary if A has orthonormal columns (rows since AA = I as well from
Problem 1.8).
Theorem 1.5.1. (Schur triangularization theorem) Let A Cnn . There is
U Un (C) such that U AU is upper triangular.
Proof. Let 1 be an eigenvalue of A with unit eigenvector x1 , i.e., Ax1 = 1 x1
with kx1 k2 = 1. Extend via Theorem 1.3.4 to an orthonormal basis {x1 , . . . , xn }
for Cn and set Q = [x1 xn ] which is unitary. Since Q Ax1 = 1 Q x1 =
1 (1, . . . , 0)T ,
1
0

A := Q AQ = .
.. A1
0
where A1 C(n1)(n1) . By induction, there is U1 Un1 (C) such that
U1 A1 U1 is upper triangular. Set

1 0 0
0

P := .
.. U 1
0
which is unitary. So
1
0

P AP = .
.. U1 A1 U1
0
is upper triangular and set U = QP .
Theorem 1.5.2. Let T End V , where V is an inner product space. Then
there is an orthonormal basis E of V such that [T ]E
E is upper triangular.
0
Proof. For any orthonormal basis E 0 = {e01 , . . . , e0n } of V let A := [T ]E E 0 . By
Theorem 1.5.1 there is U Un (C), where n = dim V , such that U AU is
upper triangular. Since Pn U is 0unitary, E = {e1 , . . . , en } is alsoEan orthonormal
0
basis, where ej = u e
i=1 ij i , j = 1, . . . , n (check!) and [I]E = U . Hence
E E E0 E0 1
[T ]E = [I]E 0 [T ]E 0 [I]E = U AU since U = U .
Lemma 1.5.3. Let A Cnn be normal and let r be fixed. Then arj = 0 for
all j 6= r if and only if air = 0 for all i 6= r.
Proof. From arj = 0 for all j 6= r and AA = A A,
n
X X
|arr |2 = |arj |2 = (AA )rr = (A A)rr = |arr |2 + |air |2 .
j=1 i6=r
So air = 0 for all i 6= r. Then apply it on the normal A .

Theorem 1.5.4. Let A Cnn . Then

1. A is normal if and only if A is unitarily similar to a diagonal matrix.
2. A is Hermitian (psd, pd) if and only if A is unitarily similar to a real
(nonnegative, positive) diagonal matrix.
3. A is unitary if and only if A is unitarily similar to a diagonal unitary
matrix.
Proof. Notice that if A is normal, Hermitian, psd, pd, unitary, so is U AU for
any U Un (C). By Theorem 1.5.1, there is U Un (C) such that U AU is
upper triangular and also normal. By Lemma 1.5.3, the result follows. The rest
follows similarly.
From Gram-Schmidt orthgonalization (Theorem 1.3.4), one has the QR de-
composition of an invertible matrix.
Theorem 1.5.5. (QR decomposition) Each invertible A Cnn can be decom-
posed as A = QR, where Q Un (C) and R is upper triangular. The diagonal
entries of R may be chosen positive; in this case the decomposition is unique.
Problems
1. Show that upper triangular in Theorem 1.5.1 may be replaced by lower
triangular.
2. Prove that A End V is unitary if and only if kAvk = kvk for all v V .
3. Show that A Cnn is psd (pd) if and only if A = B B for some (invert-
ible) matrix B. In particular B may be chosen lower or upper triangular.
4. Let V be an inner product space. Prove that (i) (Av, v) = 0 for all v V
if and only if A = 0, (ii) (Av, v) R for all v V if and only if A is
Hermitian. What happens if the underlying field C is replaced by R?
5. Show that if A is psd, then A is Hermitian. What happens if the underlying
field C is replaced by R?
6. Prove Theorem 1.5.5.
7. Prove that if T Hom (V, W ) where V and W are inner product spaces,
then T T 0 and T T 0. If T is invertible, then T T > 0 and T T > 0.

1. Apply Theorem 1.5.1 on A and take complex conjugate transpose back.
1.5. NORMAL OPERATORS AND MATRICES 15
2. A End V is unitary A A = IV So unitary A implies kAvk2 =

(Av, Av) = (A Av, v) = (v, v) = kvk2 . Conversely, kAvk = kvk for all
v V implies ((A A I)v, v) = 0 where A A I is Hermitian. Apply
Problem 5.
3. If A = B B, then x Ax = xB Bx = kBxk22 0 for all x Cn . Con-
versely if A is psd, we can define a psd A1/2 via Theorem 1.5.4. Apply QR
on A1/2 to have A1/2 = QR where Q is unitary and R is upper triangular.
Hence A = A1/2 A1/2 = (A1/2 ) A1/2 = R Q QR = R R. Set B := R.
Let A1/2 = P L where P is unitary and L is lower triangular (apply QR
decomposition on (A )1 ). Then A = L L.
4. (i) It suffices to show for Hermitian A because of Hermitian decomposition
A = H +iK where H := (A+A )/2 and K = (AA )/(2i) are Hermitian
and
(Av, v) = (Hv, v) + i(Kv, v)
and (Hv, v), (Kv, v) R. So (Av, v) = 0 for all v if and only if (Hv, v) =
(Kv, v) = 0 for all v V . Suppose A is Hermitian and (Av, v) = 0 for all
v V . Then there is an orthonormal basis of eigenvectors {v1 , . . . , vn } of
A with corresponding eigenvalues 1 , . . . , n , but all eigenvalues are zero
because i (vi , vi ) = (Avi , vi ) = 0. When C is replaced by R, we cannot
conclude that A = 0 since for any real skew symmetric matrix A Rnn ,
xT Ax = 0 (why?).
(Brice) Notice that A is psd (because (Av, v) = 0). From Problem 3 via
matrix-operator approach, A = B B for some B End V . Thus from
Problem 4.2, A = A .
(ii)
2[(Av, w) + (Aw, v)] = (A(v + w), v + w) (A(v w), v w) R
so that Im [(Av, w) + (Aw, v)] = 0 for all v, w. Similarly
2i[(Av, w) + (Aw, v)] = (A(v + iw), v + iw) (A(v iw), v iw) R
so that Re [(Av, w) + (Aw, v)] = 0 for all v, w. So
(Av, w) = (Aw, v) = (v, Aw),
i.e., A = A . When C is replaced by R, we cannot conclude that A is real

symmetric since for any real skew symmetric matrix A Rnn , xT Ax = 0.
5. Follow from Problem 4.
6. Express the columns (basis of Cn ) of A in terms of the (orthonormal)
columns of Q.
7. (T T v, v) = (T v, T v) 0 for all v V . So T T 0 and similar for
T T 0. When T is invertible, T v 6= 0 for all v 6= 0 so (T T v, v) =
(T v, T v) > 0.
1.6 Inner product and positive operators

Theorem 1.6.1. Let V be an inner product space with inner product (, ) and
let T End V . Then hu, vi := (T u, v), u, v V , defines an inner product if and
only if T is pd with respect to (, ).
Proof. Suppose hu, vi := (T u, v), u, v V , defines an inner product, where (, )

is an inner product for V . So
(T u, v) = hu, vi = hv, ui = (T v, u) = (T u, v)
for all u, v V . So T = T , i.e., self-adjoint. For any v 6= 0, 0 < hv, vi = (T v, v)

so that T is pd with respect to (, ). The other implication is trivial.
The next theorem shows that in the above manner inner products are in
one-one correspondence with pd matrices.
Theorem 1.6.2. Let (, ) and h, i be inner products of V . Then there exists a

unique T End V such that hu, vi := (T u, v), u, v V . Moreover, T is positive
definite with respect to both inner products.
Proof. Let E = {e1 , . . . , en } be an orthonormal

Pn basis of V with respect to (, ).
So each v V can be written as v = i=1 (v, ei )ei . Now for each v V defines
T End V by
Xn
T v := hv, ei iei
i=1
Clearly
n
X n
X
(T u, v) = hu, ei i(ei , v) = hu, (v, ei )ei i = hu, vi.
i=1 i=1
Since h, i is an inner product, from Theorem 1.6.1 T is pd with respect to (, ).

When v 6= 0, hT v, vi = (T 2 v, v) = (T v, T v) > 0) so that T is pd with respect to
h, i. The uniqueness follows from Problem 4.1.
Theorem 1.6.3. Let F = {f1 , . . . , fn } be a basis of V . There exists a unique

inner product (, ) on V such that F is an orthonormal basis.
Proof. Let (, ) be an inner product with orthonormal basis E = {e1 , . . . , en }.

By Problem 1.4 S End V defined by Sfi = ei is invertible. Set T := S S > 0
( and T > 0 from Problem 5.7 are with respect to (, )). So hu, vi = (T u, v) is
an inner product by Theorem 1.6.1 and f1 , . . . , fn are orthonormal with respect
to h, i. It is straightforward to show the uniqueness.
Problems
1.7. INVARIANT SUBSPACES 17
Pn
1. Let
Pn E = {e1 , . . . , en } be a basis
Pnof V . For any u = i=1 ai ei and v =
i=1 bi ei , show that (u, v) := i=1 ai bi is the unique inner product on V
so that E is an orthonormal basis.
2. Find an example that there are inner products (, ) on h, i on V and

T End V so that T is pd with respect to one but not to the other.
3. Let V be an inner product space. Show that for each A Cnn there are
u1 , . . . , un ; v1 , . . . , vn Cn such that aij = (ui , vj ) 1 i, j n.
4. Suppose that (, ) on h, i are inner products on V , T End V and let

T () and T hi be the corresponding adjoints. Show that T () and T hi are
similar.
1. Uniqueness follows from Theorem 1.6.3.
2.
3.
4. By Theorem 1.6.1 there is a pd S End V with respect to both inner

products such that hu, vi = (Su, v) for all u, v V and thus hS 1 u, vi =
(u, v). So
hu, T hi vi = hT u, vi = (ST u, v) = (u, T () S () v) = hS 1 u, T () S () vi
Since S () = S hi = S , say, and (A1 ) = (A )1 for any A End V

((A1 ) A = (AA1 ) = I by Problem 4.2). Then
hu, T hi vi = hu, (S 1 ) T () S vi = hu, (S )1 T () S vi.
Hence T hi = (S )1 T () S .
1.7 Invariant subspaces

Let W be a subspace of V . For any T Hom (V, U ), the restriction of T ,
denote by T |W , is the unique T1 Hom (W, U ) such that T1 (w) = T (w) for all
w W.
Let W be a subspace of V and T End V . Then W is said to be invariant
under T if T w W for all w W , i.e., T (W ) W . The trivial invariant
subspaces are 0 and V . Other invariant subspaces are called nontrivial or
proper invariant subspaces. If W is invariant under T , then the restriction
T |W Hom (W, V ) of T induces T |W End W (we use the same notation
T |W ).
Theorem 1.7.1. Let V be an inner product space over C and T End V . If

W is an invariant subspace under T and T , then
(a) (T |W ) = T |W .
(b) If T is normal, Hermitian, psd, pd, unitary, so is T |W .
Proof. (a) For any x, y W , T |W x = T x, T |W y = T y. By the assumption,
(T |W x, y)W = (T x, y)V = (x, T y)V = (x, T |W y)W .
So (T |W ) = T |W .
(b) Follows from Problem 7.3.
Problems
1. Prove that if W V is a subspace and T1 Hom (W, V ), then there is

T Hom (V, U ) so that T |W = T1 . Is T unique?
2. Show that the restriction on W V of the sum of S, T Hom (V, U ) is

the sum of their restrictions. How about restriction of scalar multiple of
T on W ?
3. Show that if S, T End V and the subspace W is invariant under S and

T , then (ST )|W = (S|W )(T |W ).
4. Let T End V and S End W and H Hom (V, W ) satisfy SH = HT .

Prove that Im H is invariant under S and Ker H is invariant under T .
1. Let {e1 , . . . , em } be a basis of W and extend it to E = {e1 , . . . , em , em+1 , . . . , en }

a basis of V . Define T Hom (V, U ) by T ei = T1 ei , i = 1, . . . , m and
T ej = wj where j = m + 1, . . . , n and wm+1 , . . . , wn W are arbitrary.
Clearly T |W = T1 but T is not unique.
2. (S + T )|W v = (S + T )v = Sv + T v = S|W v + T |W v for all v V . So

(S + T )|W = S|W + T |W .
3. (S|W )(T |W )v = S|W (T v) = ST v since T v W and thus (S|W )(T |W )v =

ST |W v for all v.
4. For any v Ker H, H(T v) = SHv = 0, i.e., T v Ker H so that Ker H

is an invariant under T . For any Hv Im H (v V ), S(Hv) = HT (v)
Im H.
1.8. PROJECTIONS AND DIRECT SUMS 19
1.8 Projections and direct sums

The vector space V is said to be a direct sum of the Pnsubspaces W1 , . . . , Wm
if each v V can be uniquely expressed as v = i=1 wi where wi Wi ,
i = 1, . . . , n and is denoted by V = W1 Wm . In other words, V =
W1 + + Wm and if w1 + + wm = w10 + + wm
0
where wi Wi , then
0
wi = wi for all i.
Theorem 1.8.1. V = W1 Wm if and only if V = W1 + + Wm and
Wi (W1 + +Wi + +Wm ) = 0 for all i = 1, . . . , m. Here Wi denotes deletion.
In particular, V = W1 W2 if and only if W1 + W2 = V and W1 W2 = 0.
Proof. Problem 7.
In addition, if V is an inner product space, and Wi Wj if i 6= j, i.e.,
wi wj whenever wi Wi , wj Wj , we call V an orthogonal sum of
W1 , . . . , Wm , denoted by V = W1 + +Wm .
Suppose V = W1 Wm . Each Pi End V defined by Pi v = wi ,
i = 1, . . . , m satisfies Pi2 = Pi and Im Pi = Wi . In general P End V is called
a projection if P 2 = P . Notice that P is a projection if and only if IV P is
a projection; in this case Im P = Ker (IV P ).
Theorem 1.8.2. Let V be a vector space and let P End V be a projection.
Then the eigenvalues of P are either 0 or 1 and rank P = tr P .
Proof. Let rank P = k. By Theorem 1.1.2, there are v1 , . . . , vn V such that
P v1 , . . . , P vk is a basis of Im P and {vk+1 , . . . , vn } is a basis of Ker P . From
P 2 = P , E = {P v1 , . . . , P vk , vk+1 , . . . , vn } is linearly independent (check) and
thus a basis of V . So [P ]E E = diag (1, . . . , 1, 0, . . . , 0) in which there are k ones.
So we have the desired results.
The notions direct sum and projections are closely related.
Theorem 1.8.3. Let V be a vector space and let P1 , . . . , Pm be projections
such that P1 + + Pm = IV . Then V = Im P1 Im Pm . Conversely, if
V = W1 Wm , then there are unique projections P1 , . . . , Pm such that
P1 + + Pm = IV and Im Pi = Wi , i = 1, . . . , m.
Proof. From P1 + +Pm = IV , each v V can be written as v = P1 v+ +Pm v
so that V = Im P1 + + Im Pm . By Theorem 1.8.2
m
X m
X m
X
dim V = tr IV = tr Pi = rank Pi = dim Im Pi .
i=1 i=1 i=1
So V = Im P1 Im Pm (Problem 8.1).
Conversely if V = W1 Wm , then for any v V , there is unique
factorization v = w1 + + wm , wi Wi , i = 1, . . . , m. Define Pi End V
by Pi v = wi . It is easy to see each Pi is a projection and Im Pi = Wi and
P1 + + Pm = IV . For the uniqueness, if there are projections Q1 , . . . , Qm
such that Q1 + + Qm = IV and Im Qi = Wi for all i, then for each v V ,
from the unique factorization Qi v = Pi v so that Pi = Qi for all i.
Corollary 1.8.4. If P End V is a projection, then V = Im P Ker P .

Conversely if V = W W 0 , then there is a projection P End V such that
Im P = W and Ker P = W 0 .
We call P End V an orthogonal projection if P 2 = P = P . Notice that
P is an orthogonal projection if and only if IV P is an orthogonal projection.
The notions orthogonal sum and orthogonal projections are closely related.
Theorem 1.8.5. Let V be an inner product space and let P1 , . . . , Pm be orthog-
onal projections such that P1 + + Pm = IV . Then V = Im P1 + +Im Pm .
Conversely, if V = W1 + +Wm , then there are unique orthogonal projections
P1 , . . . , Pm such that P1 + + Pm = IV and Im Pi = Wi , i = 1, . . . , m.
Proof. From Theorem 1.8.3, if P1 , . . . , Pm are orthogonal projections such that
P1 + + Pm = IV , then V = Im P1 Im Pm . It suffices to show that the
sum is indeed orthogonal. From the proof ofPTheorem 1.8.3, Pi v = vi Im Pi
m
for all i, where v V is expressed as v = i=1 vi . Thus for all u, v V , if
i 6= j, then
(Pi u, Pj v) = (Pi u, vj ) = (u, Pi vj ) = (u, Pi vj ) = (u, 0) = 0.
Conversely if V = W1 + +Wm , then from Theorem 1.8.3 there are unique

projections P1 , . . . , Pm such that P1 + + Pm = IV and Im Pi = Wi , for all
i. It remains to show that each projection
Pm Pi is an
Pmorthogonal projection, i.e.,
Pi = Pi . For u, v V , write u = i=1 ui , v = i=1 vi where ui , vi Wi for
all i. Then for all i
m
X m
X
(Pi u, v) = (ui , v) = (ui , vi ) = (ui , vi ) = ( ui , vi ) = (u, Pi v).
i=1 i=1
Corollary 1.8.6. Let V be an inner product space. If P End V is an orthog-

onal projection, then V = Im P +Ker P . Conversely if V = W +W 0 , then there
is an orthogonal projection P End V such that Im P = W and Ker P = W 0 .
Proof. Apply Theorem 1.8.5 on P and IV P .
Problems
1. Show that if V = W1 + + Wm , then V = W1 Wm if and only if
dim V = dim W1 + + dim Wm .
2. Let P1 , . . . , Pm End V be projections on V and P1 + + Pm = IV .
Show that Pi Pj = 0 whenever i 6= j.
3. Let P1 , . . . , Pm End V be projections on V . Show that P1 + + Pm is
a projection if and only if Pi Pj = 0 whenever i 6= j.
1.8. PROJECTIONS AND DIRECT SUMS 21
4. Show that if P End V is an orthogonal projection, then kP vk kvk for

all v V .
5. Prove that if P1 , . . . , Pm End V are projections on V with P1 + +

Pm = IV , then there is an inner product so that P1 , . . . , Pn are orthogonal
projections.
6. Show that if P End V is an orthogonal projection on the inner product

space and if W is P -invariant subspace of V , then P |W End W is an
orthogonal projection.
1.
2. By Theorem 1.8.3, V = Im P1 Im Pm so that for any v V ,

Pi Pj v = Pi vj = 0 if i 6= j.
3. P1 + + Pm is a projection means
X X X X
Pi + Pi Pj = Pi2 + Pi Pj = (P1 + + Pm )2 = P1 + + Pm .
i i6=j i i6=j
So if Pi Pj = 0 for i 6= j, then P1 + + Pm is a projection. Conversely, if

P1 + + Pm is a projection, so is IV (P1 + + Pm ).
(Roy) Suppose that P := P1 + +Pm is a projection. Then the restriction
of P on its image P (V ) is the identity operator and Pi |P (V ) is still a
projection for each i = 1, . . . , m. According to Theorem 1.8.3, P (V ) =
m m
i=1 Im Pi |P (V ) = i=1 Im Pi , which implies that Pi Pj = 0 whenever i 6= j.
4. I P is also an orthogonal projection. Write v = P v + (I P )v so that

kvk2 = kP vk2 + k(IP )vk2 kP vk2 .
5. By Theorem 1.8.3, V = Im P1 Im Pm . Let {ei1 , . . . , eini } be a

basis of Im Pi for all i. Then there is an inner product, by Theorem 1.6.3
so that the basis {e11 , . . . , e1n1 , . . . , em1 , . . . , emnm } is orthonormal. Thus
P1 , . . . , Pm are orthogonal projections by Theorem 1.8.5.
6. Notice that P |2W = P 2 |W = P |W since P 2 = P and from Theorem 1.7.1

(P |W )= P |W = P |W since P = P .
7.
1.9 Dual spaces and Cartesian products

The space V = Hom (V, C) is called the dual space of V . Elements in V are
called (linear) functionals.
Theorem 1.9.1. Let E = {e1 , . . . , en } be a basis of V . Define e1 , . . . , en V

by
ei (ej ) = ij , i, j = 1, . . . , n.
Then E = {e1 , . . . , en } is a basis of V .
Pn
Proof. Notice that each f V can be written as f P = i=1 f (ei )ei since both
n
sides
Pn take the same values on e1 , . . . , en . Now if i=1 ai ei = 0, then 0 =

i=1 ai ei (ej ) = aj , j = 1, . . . , n. So e1 , . . . , en are linearly independent.
The basis E is called a basis of V dual to E, or simply dual basis.

If T Hom (V, W ), then its dual map (or transpose) T Hom (W , V )
is defined by
T () = T, W .
The functional T () is in V , and is called the pullback of along T . Imme-
diately the following identity holds for all W and v V :
hT (), vi = h, T (v)i
where the bracket hf, vi := f (v) is the duality pairing of V with its dual space,
and that on the right is the duality pairing of W with its dual. This identity
characterizes the dual map T , and is formally similar to the definition of the
adjoint (if V and W are inner product spaces). We remark that we use the same
notation for adjoint and dual map of T Hom (V, W ). But it should be clear
in the context since adjoint requires inner products but dual map does not.
Theorem 1.9.2. The map : Hom (V, W ) Hom (W , V ) defined by (T ) =

T is an isomorphism.
Proof. Problem 10.
When V has an inner product, linear functionals have nice representation.
Theorem 1.9.3. (Riesz) Let V be an inner product space. For each f V ,

there is a unique u V such that f (v) = (v, u) for all v V . Hence the map
: V V defined by (u) = (, u) is an isomorphism.
Proof. Let E = {e1 , . . . , en } be an orthonormal basis of V . Then for all v V ,

by Theorem 1.3.3
n
X n
X n
X
f (v) = f ( (v, ei )ei ) = f (ei )(v, ei ) = (v, f (ei )ei )
i=1 i=1 i=1
1.9. DUAL SPACES AND CARTESIAN PRODUCTS 23
Pn
so that u := i=1 f (ei )ei . Uniqueness follows from the positive definiteness of
the inner product: If (v, u0 ) = (v, u) for all v V , then (v, u u0 ) = 0 for all v.
Pick v = u u0 to have u = u0 .
The map : V V defined by (u) = (, u) is clearly a linear map bijective
from the previous statement.
Let V1 , . . . , Vm be m vector spaces over C.
m
i Vi = V1 Vm = {(v1 , . . . , vm ) : vi Vi , i = 1, . . . , m}
is called the Cartesian product of V1 , . . . , Vm . It is a vector space under the

natural addition and scalar multiplication.
Remark: If V is not finite-dimensional but has a basis {e : A} (axiom of
choice is needed) where A is the (infinite) index set, then the same construction
as in the finite-dimensional case (Theorem 1.9.1) yields linearly independent
elements {e : A} in V , but they will not form a basis.
Example (for Alexs question): The space R , whose elements are those
sequences of real numbers which have only finitely many non-zero entries, has
a basis {ei : i = 1, 2, . . . , } where ei = (0, . . . , 0, 1, 0, . . . ) in which the only
nonzero entry 1 is at the ith position. The dual space of R is R , the space of
all sequences of real
Pnumbers and such a sequence (an ) is applied to an element
(xn ) R to give n an xn , which is a finite sum because there are only finitely
many nonzero xn . The dimension of R is countably infinite but R does not
have a countable basis.
Problems
1. Show that Theorem 1.9.3 remains true if the inner product is replaced by a
nondegenerate bilinear form B(, ) on a real vector space. Nondegeneracy
means that B(u, v) = 0 for all v V implies that u = 0.
2. Let {e1 , . . . , en } be an orthonormal basis of the inner product space V .

Show that {f1 , . . . , fn } is a dual basis if and only if fj (v) = (v, ei ) for all
v V and j = 1, . . . , n.
3. Let {f1 , . . . , fn } be a basis of V . Show that if v V such that fj (v) = 0

for all j = 1, . . . , n, then v = 0.
4. Let {f1 , . . . , f } be a basis of V . Prove that there is a basis E = {e1 , . . . , en }

of V such that fi (ej ) = ij , i, j = 1, . . . , n.
5. Let E = {e1 , . . . , en } and F = {f1 , . . . , fn } be two bases of V and let E

E
and F be their dual bases. If [IV ]F E = P and [IV ]F = Q, prove that
Q = (P 1 )T .
6. Show that if S Hom (W, U ) and T Hom (V, W ), then (ST ) = T S .

7. Show that : V V defined by (v)() = (v) for all v V and

V V , is an (canonical) isomorphism.
8. Suppose E and F are bases for V and W and let E and F be their dual
bases. For any T Hom (V, W ), what is the relation between [T ]F
E and

[T ]E
F ?
9. Show that dim(V1 Vm ) = dim V1 + + dim Vm .

1. Let B(, ) be a nondegenerate bilinear form on a real vector space V .
Define : V V by (v) = B(v, ). Clearly is linear and we need to
show that is surjective. Since dim V = dim V , it suffices to show that
is injective (and thus bijective). Let v Ker , i.e., B(v, w) = 0 for all
w V . By the nondegeneracy of B(, ), v = 0.
2. Notice
Pn that (ei , ej ) = ij .P If F is a dualPbasis, then for each v =
n n
i=1 (v, ei )ei , fj (v) = fj ( i=1 (v, ei )ei ) = i=1 (v, ei )fj (ei ) = (v, ej ).
On the other hand, if fj (v) = (v, ej ) for all v, then fj (ei ) = (ei , ej ) = ij ,
i.e., F is a dual basis.
3. The assumption implies that f (v) = 0 for all f V . If v 6= 0, extends it
to basis E = {v, v2 , . . . , vn } of V . Define g V by g(v) = 1 and g(vi ) = 0
for all i = 2, . . . , n.
4. Introduce an inner product on V so that by Theorem 1.9.3 we determine
u1 , . . . , un via fj (v) = (v, uj ). By Theorem 1.6.3 and Theorem 1.6.1 there
is a pd T End V such that (T ui , uj ) = ij . Set ei = T ui for all i.
Uniqueness is clear.
Another approach: Since {f1 , . . . , fn } is a basis, each fi 6= 0 so that
dim Ker fi = n 1. So pick vi j6=i Ker fj 6= 0 (why?) such that
fi (vi ) = 1. Then fj (vi ) = 0 for all j 6= 0.
5.
6. For any U , (T S ) = T ( S) = ( S) T = (ST ) = (ST ) .
So (ST ) = T S .
7. Similar to Problem 10. Notice that (ei ) = e i for all i. It is canonical
because of e 7 e .
Pn P
8. Let A = [T ]F E , i.e., T ej = i=1 aij fi . Write v = i i ei so that
X X X X
(T fj )v = fj (T v) = fj ( i T ei ) = fj ( i aki fk ) = i aki fj (fk ) = i aji .
i i,k i,k i
1.10. NOTATIONS 25
On the other hand

X X X X
( aji ei )(v) = aji ei ( k ek ) = i aji .
i i k i
P E
So T fj =
i aji ei , i.e., [T ] = AT , i.e., ([T ]F
F E)
T
= [T ]E
F . This

explains why T is also called the transpose of T .
(Roy) Suppose E = {e1 , . . . , en } and F = {f1 , . . . , fm }. Let Pm [T ]F
E =
E
(aij ) CmnPand [T ]F = (bpq ) Cnm . By definition, T ej = i=1 aij fi
n
and T fq = p=1 bpq ep . Since ej (ei ) = ij and fj (fi ) = ij , the definition

(T fj )ei = fj (T ei ) implies that bij = aji . Thus [T ]E F T
F = ([T ]E ) .
9. If Ei = {ei1 , . . . , eini } is a basis of Vi , i = 1, . . . , m, then
E = {(e11 , . . . , em1 ), . . . , (e1n1 , . . . , emnm )}
is a basis of V1 Vm (check!).
10. is linear since for any scalars , , (S + T ) = (S + T ) =

S + T = S + T = (S + T ). It is injective since if
T = 0, then T = T = 0 for all V . By Problem 3, T = 0. Thus
is an isomorphism.
1.10 Notations
Denote by Sm the symmetric group on {1, . . . , m} which is the group of all
bijections on {1, . . . , m}. Each Sm is represented by

1, ..., m
= .
(1), . . . , (m)
The sign function on Sm is

(
1 if is even
() =
1 if is odd
where Sm . Let , , denote finite sequences whose components are natural

numbers
(n1 , . . . , nm ) := { : = ((1), . . . , (m)), 1 (i) ni , i = 1, . . . , m}

m,n := { : = ((1), . . . , (m)), 1 (i) n, i = 1, . . . , m}
Gm,n := { m,n : (1) (m)}
Dm,n := { m,n : (i) 6= (j) whenever i 6= j}
Qm,n := { m,n : (1) < < (m)}.
For Qm,n and Sm , define

:= (((1)), . . . , (m))).
Then Dm,n and
Dm,n = { : Qm.n , Sm } = Qm,n Sm . (1.1)
Suppose n > m. For each Qm,n , denote by 0 the complementary
sequence of , i.e., 0 Qnm,n , the components of and 0 are 1, . . . , n, and

1, ..., m, m + 1, . . . , n
= Sn .
(1), . . . , (m), 0 (1), . . . , 0 (n m)
It is known that
() = (1)s()+m(m+1)/2 , (1.2)
where s() := (1)+ +(m). Similarly for Qm,n , Sm and Snm ,

1, ..., m, m + 1, . . . , n
= Sn
(1), . . . , (m), 0 (1), . . . , 0 (n m)
and
() = ()()(1)s()+m(m+1)/2 , (1.3)
Moreover

1, ..., m, m + 1, . . . , n
Sn = { = :
(1), . . . , (m), 0 (1), . . . , 0 (n m)
Qm,n , Sm , Snm }. (1.4)
Let A Cnk . For any 1 m n, 1 l k, Qm,n and Ql,k . Let
A[|] denote the submatrix of A by taking the rows (1), . . . , (m) of A and the
columns (1), . . . , (l). So the (i, j) entry of A[|] is a(i)(j) . The submatrix
of A complementary to A[|] is denoted by A(|) := A[0 | 0 ] C(nm)(kl) .
Similarly we define A(|] := A[0 |] C(nm)k and A[|) := A[| 0 ]
Cn(kl) .
Recall that if A Cnn , the determinant function is given by
X n
Y
det A = () ai(i) . (1.5)
Sn i=1
It is easy to deduce that

X n
Y
det A = ()() a(i),(i) . (1.6)
Sn i=1
Let A Cnk . For any 1 m min{n, k}, Qm,n and Qm,k ,

X m
Y
det A[|] = () a(i),(i) . (1.7)
Sm i=1
1.10. NOTATIONS 27
For any Sn , we have

det A[|] = det A[|] = () det A[|]. (1.8)
When m,n , m,k , A[|] Cmm whose (i, j) entry is a(i)(j) .
However if m,n and 6 Dm,n , then A[|] has two identical rows so that
det A[|] = 0, m,n \ Dm,n . (1.9)
Theorem 1.10.1. (Cauchy-Binet) Let A Crn , B Cnl and C = AB.
Then for any 1 m min{r, n, l} and Qm,r and Qm,l ,
X
det C[|] = det A[|] det B[|].
Qm,n
Proof.
X m
Y
det C[|] = () c(i),(i)
Sm i=1
X Ym X n
= () a(i),j bj,(i)
Sm i=1 j=1
X m
X Y
= () a(i),(i) b(i),(i)
Sm m,n i=1
m
X Y X
= a(i),(i) ()b(i),(i)
m,n i=1 Sm
m
X X
= a(i),(i) det B[|]
m,n i=1
m
X Y
= a(i),(i) det B[|]
Dm,n i=1
X m
X Y
= a(i),(i) det B[|]
Qm,n Sm i=1
X X m
Y
= ( () a(i),(i) ) det B[|]
Qm,n Sm i=1
X
= det A[|] det B[|].
Qm,n
Theorem 1.10.2. (Laplace) Let A Cnn , 1 m n and Qm,n . Then

X
det A = det A[|](1)s()+s() det A(|).
Qm,n
Proof. The right side of the formula is

X
(1)s() (1)s() det A[|] det A[0 | 0 ]
Qm,n
X X X m
Y nm
Y
s() s()
= (1) ()()(1) a(i)(i) a0 (j)0 (j)
Qm,n Sm Snm i=1 j=1
X m
Y n
Y
= (1)s() ()(1)m(m+1)/2 a(i)(i) a0 (im)(i)
Sn i=1 i=m+1
X n
Y
= ()() a(i)(i)
Sn i=1
= det A.
Theorem 1.10.3. Let U Un (C). Then for , Qm,n with 1 m n,
det U det U [|] = (1)s()+s() det U (|). (1.10)
In particular
| det U [|]| = | det U (|)|. (1.11)
Proof. Because det I[|] = and U [|] = U [|]T , apply Cauchy-Binet
theorem on In = U U to yield
X
= det U [|] det U [|]. (1.12)
Qm,n
Using Laplace theorem, we have

X
det U = det U [|](1)s()+s() det U (|).
Qm,n
Multiplying both sides by U [|] and sum over Sm , the left side becomes
X
det U det U [|] = det U det U [|]
Qm,n
and the right side becomes

X X
det U [|] det U [|](1)s()+s() det U (|)
Qm,n Qm,n
X
= (1)s()+s() det U (|)
Qm,n
= (1)s()+s() det U (|)
Since | det U | = 1, we have the desired result.

1.10. NOTATIONS 29
The permanent function of A is

n
X Y
per A = ai(i)
Sn i=1
which is also known as positive determinant.
Problems
1. Prove equation (1.2).
Qm Pn P Qm
2. Prove that i=1 j=1 aij = m,n i=1 ai(i) .
3. Prove a general Laplace expansion theorem: Let A Cnn , , Qm,n ,
and 1 m n.
X
det A = det A[|](1)s()+s() det A(|).
Qm,n
4. Show that
m
Y
m n+m1
|(n1 , . . . , nm )| = ni , |m,n | = n , |Gm,n | = ,
i=1
m

n n
|Dm,n | = m!, |Qm,n | = .
m m

1.
Qm Pn P Qm
2. (Roy) If A = (aij ) Cmn , the left side of i=1 j=1 aij = m,n i=1 ai(i)
is the product of row sums of A. Fix n and use induction on m. If m = 1,
then both sides are equal to the sum of all entries of A. Suppose that the
identity holds for m 1. We have

X Y m X m1 Y Xn
ai(i) = ai(i) amj
m,n i=1 m1,n i=1 j=1

m1
YX n Xn
= aij amj by induction hypothesis
i=1 j=1 j=1
m X
Y n
= aij
i=1 j=1
3.
4. The
Qm ith slot of each sequence inm(n1 , . . . , nm ) has ni choices. So |(n Q1m, . . . , nm )| =
i=1 n i and thus |m,n | = n follows from |(n 1 , . . . , nm )| = i=1 ni
with ni = n for all i.

|Gm,n | = n+m1 m . Now |Qm,n | is the number of picking m distinct num-
n
bers
n from 1, . . . , n. So |Qm,n | = m is clear and |Dm,n | = |Qm,n |m! =
m m!
Chapter 2
Group representation
theory
2.1 Symmetric groups

Let Sn be the symmetric group on {1, . . . , n}. A matrix P Cnn is said to be
a permutation matrix if
pi,(j) = i,(j) , i, j = 1, . . . , n
where Sn and we denote by P () for P because of its association with

. Denote by Pn the set of all permutation matrices in Cnn . Notice that
: Sn Pn ( 7 P ()) is an isomorphism.
Theorem 2.1.1. (Cayley) Each finite group G with n elements is isomorphic
to a subgroup of Sn .
Proof. Let Sym (G) denotes the group of all bijections of G. For each G,
define the left translation l : G G by l (x) = x for all x G. It is easy
to show that l = l l for all , G so that l : G Sym (G) is a group
homomorphism. The map l is injective since l = l implies that x = x for
any x G and thus = . So G is isomorphic to some subgroup of Sym (G)
and thus to some subgroup of Sn .
The following is another proof in matrix terms. Let G = {1 , . . . , n }. Define
the regular representation Q : G GLn (C):
Q()ij := (i ,j ) GLn (C).
It is easy to see that Q() is a permutation matrix and Q : 7 Q() is injective.

So it suffices to show that Q is a homomorphism. Now
n
X n
X
(Q()Q())ij = Q()ik Q()kj = i ,k k ,j = i ,j = Q()ij .
k=1 k=1
31
32 CHAPTER 2. GROUP REPRESENTATION THEORY
So we view each finite group as a subgroup of Sn for some n.

The element Sn is called a cycle of length k (1 k n), if there exist
1 i1 , . . . , ik n such that
(it ) = it+1 , t = 1, . . . , k 1
(ik ) = i1
(i) = i, i 6 {i1 , . . . , i + k}
and we write = (i1 , . . . , ik ). Two cycles (i1 , . . . , ik ) and (j1 , . . . , jm ) are said
to be disjoint if there is no intersection between the two set {i1 , . . . , ik } and
{j1 , . . . , jm }. From the definition, we may express the cycle as
(i1 , (i1 ), . . . , k1 (i1 )).
For any Sn there is 1 k n such that k (i) = i and let k be the

smallest such integer. Then (i (i) k1 (i)) is a cycle but is not necessarily
equal to . The following theorem asserts that can be reconstructed from
these cycles.
Theorem 2.1.2. Each element Sn can be written as a product of disjoint
cycles. The decomposition is unique up to permutation of the cycles.
Proof. Let i n be a positive integer. Then there is a smallest positive integer
r such that r (i) = i. If r = n, then the cycle (i (i) r1 (i)) is . If r < n,
then there is positive integer j n such that j 6 {i, (i), . . . , r1 (i)}. Then
there is a smallest positive integer s such that s (j) = j. Clearly the two cycles
(i (i) r1 (i)) and (j (j) s1 (j)) are disjoint. Continue the process
we have
= (i (i) r1 (i))(j (j) s1 (j)) (k (k) t1 (k))
For example,

1 2 3 4 5 6 7
= 1 2 5 3 4 7 6 = 1 2 5 3 4 7 .
2 5 1 7 3 6 4
So c() = 3 (which include the cycle of length 1)

1 2 3 4 5
= 1 2 5 3 4 = 3 4 1 2 5 = 3 4 5 1 2 .
2 5 4 3 1
So c() = 2.
Cycles of length 2, i.e., (i j), are called transpositions.
Theorem 2.1.3. Each element Sn can be written (not unique) as a product
of transposition.
2.1. SYMMETRIC GROUPS 33
Proof. Use Theorem 2.1.2

(i1 ik ) = (i1 i2 )(i2 i3 ) (ik1 ik ) = (i1 ik )(i1 ik1 ) (i1 i2 ). (2.1)
Though the decomposition is not unique, the parity (even number or odd
number) of transpositions in each Sn is unique (see Merris p.56-57).
Problems
1. Prove that each element of Sn is a product of transpositions. (Hint:
(i1 i2 ik ) = (i1 i2 )(i2 i3 ) (ik1 ik ) = (i1 ik )(i1 ik1 ) (i1 i2 )).
2. Show that Sn is generated by the transposition (12), (23), . . . , (n 1, n)
(Hint: express each tranposition as a product of (12), (23), . . . , (n 1, n)).
3. Express the following permutations as a product of nonintersecting cycles:
(a) (a b)(a i1 . . . ir b j1 . . . jr ).
(b) (a b)(a i1 . . . ir b j1 . . . jr )(a b).
(c) (a b)(a i1 . . . ir ).
(d) (a b)(a i1 . . . ir )(a b).
4. Express = (24)(1234)(34)(356)(56) S7 as a product of disjoint cycles
and find c().

1. Use Theorem 2.1.2 and the hint.
2. (i, i + 2) = (i + 1, i + 2)(i, i + 1)(i + 1, i + 2) and each (ij) is obtained by
conjugating (i, i + 1) by the product (j, j + 1) (i + 1, i + 2) where i < j.
So Sn is generated by (12), (23), . . . , (n 1, n).
(Zach) (ij) = (1i)(1j)(1i). By Problem 1, every cycle is a product of
transpositions. So Sn is generated by (12), (13), . . . , (1n). Now (1k) =
(1, k 1)(k 1, k)(1, k 1) for all k > 1.
3. (a) (a b)(a i1 . . .
ir b j1 . . . jr ) = (a i1 . . . ir )(b j1 . . . js ).
(b) (a b)(a i1 . . .
ir b j1 . . . js )(a b) = (a i1 . . . ir )(b j1 . . . js )(a b) = (a j1 . . . js b i1 . . . ir ).
(c) (a b)(a i1 . . .
ir ) = (a i1 . . . ir b).
(d) (a b)(a i1 . . .
ir )(a b) = (b i1 . . . ir ).

1 2 3 4 5 6 7
4. = (24)(1234)(34)(356)(56) = = (14235)(6)(7).
4 3 5 2 1 6 7
So c() = 3.
2.2 Representation theory

Let V be an n-dimensional complex vector space and GL(V ) be the group of
invertible operators T End V . Clearly the group of n n complex invertible
matrices GLn (C) is isomorphic to GL(V ). Let G be a finite group. A (linear)
representation of G of degree n is a homomorphism T : G GL(V ), i.e.,
T (1 2 ) = T (1 )T (2 ), 1 , 2 G.
Then T (e) = I where e G is the identity element (reason: T (e) = T (e2 ) =

T (e)T (e)) and T ( 1 ) = (T ())1 . The dimension of V is called the degree of
the representation T . The representation T is called faithful (to the group) if
it is injective, i.e., G is isomorphic to T (G) < GL(V ).
Similarly we have the matrix representation A : G GLn (C), i.e.,
A(1 2 ) = A(1 )A(2 ), 1 , 2 G.
So A(e) = A and A( 1 ) = (A())1 . Clearly if P GLn (C), then B() :=

P 1 A()P , G, yields a matrix representation B : G GLn (C). We call
the representations A and B isomorphic.
Let E be a basis of V and let T : G GL(V ) be a representation. Let
A() := [T ()]E
E GLn (C) where n = dim V . Then A : G GLn (C) is a
matrix representation since
A(1 2 ) = [T (1 2 )]E E E E
E = [T (1 )T (2 )]E = [T (1 )]E [T (2 )]E = A(1 )A(2 ).
Under another basis F , we have a matrix representation B() := [T ()]F

F . Then
B() = [T ()]F F E E
F = [I]E [T ()]E [I]F = P
1
A()P, G,
where P := [I]E F , i.e., A and B are isomorphic. In other words, the matrix
representations with respect to different bases of a linear representation are
isomorphic. Similarly the representations S : G GL(V ) and T : G : GL(W )
are said to be isomorphic if there is an invertible L Hom (V, W ) such that
T () = L1 S()L for all G. In other words, isomorphic representations
have the same matrix representation by choosing appropriate bases accordingly.
Given any representation T : G GL(V ) with dim V = n, a choice of basis
in V identifies with Cn (i.e., isomorphic to via the coordinate vector) and call
this isomorphism L. Thus we have a new representation S : G GLn (C)
which is isomorphic to T . So any representation of G of degree n is isomorphic
to a matrix representation on Cn . The vector space representation does not
lead to new representation, but it is a basis free approach; while at the same
time the matrix representation may give us some concrete computation edge.
So we often use both whenever it is convenient to us.
The basic problem of representation theory is to classify all representations
of G, up to isomorphism. We will see that character theory is a key player.
Example 2.2.1. (a) The principal representation: 7 1 for all G,
is of degree 1.
2.2. REPRESENTATION THEORY 35
(b) The alternating representation on Sn , 7 (), is of degree 1.
(c) 7 In is a representation of degree n.
(d) The regular representation Q : G GLn (C), where n = |G|, is given

by
Q()ij = i ,j .
Here is another version: let V be an n-dimensional complex vector space
with a (ordered) basis indexed by G, i.e., E = {f1 , . . . , fn } where G =
{1 , . . . , n }. For each G, define T () by T ()f = f . So the
images f = T ()fe , where G, form a basis of V . Then Q = [T ()]F F
since the jth column of Q is ei where j = i , i.e., Q()ij = i ,j .
(e) S3 has 6 elements: e, (23), (12), (123), (132), (13). Define T : S3 GL2 (C)
by setting their images accordingly as

1 1 0 1 0 1 1 1 1 0
I2 , , , , ,
0 1 1 0 1 1 1 0 1 1
which is a representation of degree 2.
The representation A : G GLn (C) is said to be reducible if there is

P GLn (C) such that

A1 () 0
P 1 A()P = , for all G, (2.2)
C() A2 ()
where A1 () GLm (C) and A2 () GLnm (C) with 1 m < n. In addition if

C() = 0 for all G, then A is said to be completely reducible. Moreover,
A is said to be irreducible if A is not reducible. It is easy to see that A
is irreducible if and only if any representation isomorphic to A is irreducible.
Similar notions are defined for representation T : G GL(V ), for example, T
is reducible if there is a basis E for V such that [T ()]EE is of the form (2.2),
etc.
Notice that if A : G GLn (C) is reducible, i.e., (2.2) holds, then

A1 () 0 A1 () 0 A1 () 0
=
C() A2 () C() A2 () C() A2 ()

A1 ()A1 () 0
= .(2.3)
C()A1 () + A2 ()C() A2 ()A2 ()
So A1 () = A1 ()A1 () and A2 () = A2 ()A1 (). So A1 and A2 are lower

degree representations of G, known as sub-representations. If A is completely
reducible, it is determined by A1 and A2 .
From definition all linear representations are irreducible. So Example 2.2.1(a),
(b) are irreducible; (c) is reducible if and only if n > 1; (d) is reducible if n > 1
(we will see, Problem 1 when n = 3); (e) Problem 2.
Theorem 2.2.2. The representation T : G GL(V ) is reducible if and only

if there is a nontrivial subspace W invariant under T () for all G, i.e.,
T ()W W (indeed T ()W = W ) for all G.
Proof. If T is reducible, then there is a basis E = {e1 , . . . , em , em+1 , . . . , en }
such that
E A1 () 0
[T ()]E = , for all G,
C() A2 ()
where A1 () GLm (C) and A2 () GLnm (C). Clearly W := hem+1 , . . . , en i
is the desired invariant subspace.
Conversely if W is nontrivial and invariant under T () for all G, then
extend the basis F of W to a basis E for V in which F is the latter part of E.
Then [T ()]EE is the desired reducible representation.
So T |W : G GL(W ) is a representation of G and is called a sub-representation

of T , defined by restriction:
T |W () = T ()|W .
The following theorem asserts that reducibility and completely reducibility co-
incide for finite group representations.
Theorem 2.2.3. (Maschke) Let A be a matrix representation of the finite group
G. If A is reducible, then A is completely reducible.
Proof. Since A is reducible, we may assume that

A1 () 0
A() = , for all G,
C() A2 ()
where A1 () GLm (C) and A2 () GLnm (C). Let

Im 0 1 Im 0
P := , so that P = ,
Q Inm Q Inm
where Q is to be determined. Then

1 Im 0 A1 () 0 Im 0
P A()P =
Q Inm C() A2 () Q Inm

A1 () 0
= . (2.4)
C() QA1 () + A2 ()Q A2 ()
So we want to find Q such that C() QA1 () + A2 ()Q = 0 for all G.

From (2.3)
C() = C()A1 () + A2 ()C()
so that
C() = C()A1 ()1 A2 ()C()A1 ()1

= C()A1 ()1 A1 () A2 ()C()A1 ()1 .
2.2. REPRESENTATION THEORY 37
Sum over to have

X X
|G|C() = ( C()A1 ()1 )A1 () A2 ()( C()A1 ()1 )
G G
X X
1
= ( C()A1 () )A1 () A2 ()( C()A1 ()1 )
G G
1
P
Set Q := |G| ( G C()A1 ()1 ). Then
C() QA1 () + A2 ()Q = 0, for all G.
Substitute into (2.4) to have the desired result.
Theorem 2.2.4. Every matrix representation A of the finite group G is a

direct sum of irreducible representations, i.e., there is P GLn (C) such that
P 1 A()P = A1 () A2 () Ak () for all G, where A1 , . . . , Ak are
irreducible representations.
Proof. Use Theorem 2.2.3 with induction.
So the study of representations of G, up to isomorphism, is reduced to the
study of irreducible representations.
A representation T : G GL(V ) is unitary if V has been equipped with
an inner product such that T () are unitary for all G. It is unitarisable
if it can be equipped with such an inner product. For example, the regular
representation Q : G GLn (C) (n = |G|) of a finite group is unitary (pick the
standard inner product).
Theorem 2.2.5. Every representation T : G GL(V ) is unitarisable.
Proof. Problem 3 unitary trick.
Problems
1. Prove that the degree 3 representation of S3 : 7 (i(j) ) GL3 (C) is
reducible (Hint: find a 1-dimensional invariant subspace).
2. Prove that the degree 2 representation in Example 2.2.1(e) is irreducible.
3. Let T : G GLn (C) be a representation of a finite group G. Prove that
there is an inner product such that T () is unitary for all G, i.e.,
for any u, v V , G we have (T ()u, T ()v) = (u, v). (Hint: Let
h, iPbe any arbitrary inner product. Then consider it average (u, v) :=
1
|G| G hT ()u, T ()vi).
4. Prove that every matrix representation of a finite group is isomorphic to

a unitary matrix representation.
5. Write out the regular representation of the cyclic group C3 = {1, , 2 }

where is the cubic primitive root of unity. Is it reducible?
6. Prove Maschke Theorem using Theorem 2.2.5 (Hint: If W is an invariant
subspace for all A(), so is W ).

1. The representation T is given as

1 1
e 7 I3 , (12) 7 1 , (13) 7 1 ,
1 1

1 1 1
(23) 7 1 , (123) 7 1 , (132) 7 1
1 1 1
The subspace W := he1 + e2 + e3 i is invariant under all permutations in
GL3 (C). By Theorem 2.2.2 T : S3 GL3 (C) is reducible.
Remark: Using the same argument regular representation is always re-
ducible if |G| > 1 since the subspace spanned by e1 + + en , where
n := |G|, is a nontrivial invariant subspace.
2. If the representation were reducible, there would exist a nontrivial in-
variant subspace, i.e., 1-dimensional W , by Theorem 2.2.2. Let W =
hc1 e1 + c2 e2 i and c1 e1 + c2 e2 6= 0. Then

0 1
(c1 e1 + c2 e2 ) = c2 e1 + c1 e2 W
1 0
i.e., c2 = c1 and c1 = c2 . Hence c2 = 2 c2 and c1 = 2 = c1 . So 2 = 1,

i.e., c2 = c1 . From

1 1
(c1 e1 + c2 e2 ) = (c1 c2 )e1 c2 e2 W,
0 1
c1 c2 = c1 and c2 = c2 . If c2 6= 0, then = 1 but then 2c1 = c2 ,
which is impossible.
(Roy) The matrices corresponding to (23), (12), (13) are

1 1 0 1 1 0
, ,
0 1 1 0 1 1
with eigenvalues 1, 1. If the representation were reducible, i.e., com-
pletely reducible, then all the matrices are simultaneously similar to diag-
onal matrices, in particular the above three matrices. Two would be the
same, contradiction.
2.3. IRREDUCIBLE REPRESENTATIONS 39
3. (a) It is straight forward computation to verify that (, ) is an inner

product.
1
P 1
P
(u, v) = |G| G hT ()u, T ()vi = |G| G hT ()v, T ()ui = (v, u).
(b)
(1 u1 + 2 u2 , v)
1 X
= hT ()(1 u1 + 2 u2 ), T ()vi
|G|
G
1 X
= h1 T ()u1 + 2 T ()u2 T ()vi
|G|
G
1 X 1 X
= 1 hT ()u1 , T ()vi + 2 hT ()u2 , T ()vi
|G| |G|
G G
= 1 (u1 , v) + 2 (u2 , v)
1
P
(c) (v, v) = |G| G hT ()v, T ()vi 0 since each summand is non-
negative. Moreover (v, v) = 0 if and only if T ()v = 0 for all .
Since T () GL(V ), v = 0.
For all u, v V
1 X
(T ()u, T ()v) := hT ()T ()u, T ()T ()vi
|G|
G
1 X
= hT ()u, T ()vi
|G|
G
1 X
= hT ()u, T ()vi = (u, v)
|G|
G
4. From Problem 3.

1 1
5. e 7 I3 , 7 1 , 2 7 1. It is reducible using the
1 1
argument of Problem 1 (or observe that C3 is abelian so all irreducible
representations must be of degree 1 from Problem 3.1).
6. With respect to some inner product, each A() is unitary by Theorem
2.2.5, if A : G GLn (V ) is reducible, i.e., there is a nontrivial invariant
subspace W . For each v W , we have (A()v, A()w) = (v, w) = 0 for
all w W . But A()W = W so that (A()v, W ) = {0}, i.e., A()W
W for all A. So W is invariant under A() for all G.
2.3 Irreducible representations

Theorem 2.3.1. (Schurs lemma) Let T : G GL(V ) and S : G GL(W ) be
two irreducible representations of G. If there is nonzero L Hom (V, W ) such
that L T () = S() L for all G, then dim V = dim W and L is invertible;

thus S and T are isomorphic.
Proof. Recall
Im L = {Lv : v V } W, Ker L = {v V : Lv = 0} V.
From L 6= 0 we have Im L 6= 0 and Ker L 6= V . For any v V , from S()Lv =

LT ()v Im L, Im L is S()-invariant subspace of W for all G. By Theorem
2.2.2 Im L must be trivial, i.e., Im L = W . Similarly Ker L is T ()-invariant
subspace of V so that Ker L = 0. Hence L is invertible; thus S and T are
isomorphic.
The following is simply a matrix version of Theorem 2.3.1.
Corollary 2.3.2. Let A : G GLn (C) and B : G GLm (C) be two ir-
reducible representations of G. If there is nonzero M Cmn such that
M A() = B()M for all G, then n = m and M is invertible; thus A
and B are isomorphic.
The special case A = B in Corollary 2.3.2 provides more information on M .
Corollary 2.3.3. Let A : G GLn (C) be an irreducible representation of G.
If there is M Cnn such that M A() = A()M for all G, then M = cIn .
Proof. Let c be an eigenvalue of M . So M cIn is not invertible. Clearly
(M cIn )A() = A()(M cIn ) for all G. Then apply Corollary 2.3.2.
There are important orthogonal relations among the entries of an irreducible
matrix representation and between matrix representations.
Theorem 2.3.4. Let A() = (aij ()) GLn (C) be a degree n irreducible
matrix representation of G. Then
X |G|
ais ( 1 )atj () = ij st . (2.5)
n
G
If B() = (bij ()) is an irreducible degree m matrix representation of G and

not isomorphic to A, then
X
ais ( 1 )btj () = 0. (2.6)
G
Proof. For any M Cnm , let

X
(M ) := A( 1 )M B().
G
For any G,
X X
A()(M ) = A( 1 )M B() = A( 1 )M B( ) = (M )B().
G G
2.3. IRREDUCIBLE REPRESENTATIONS 41
If A and B are not isomorphic, then from Corollary 2.3.2, (M ) = 0; if A = B,

then from Corollary 2.3.3 we have (M ) = cM In .
Pick M = Est which is the matrix with 1 as its (s, t)-entry and zeros else-
where. Computing the (i, j)-entry of (Est ):
X
(Est )ij = (A( 1 )Est B())ij
G
n X
XX m
= A( 1 )ip (Est )pq B()qj
G p=1 q=1
XX n X m
= aip ( 1 )sp tq bqj ()
G p=1 q=1
X
= ais ( 1 )btj ().
G
If A and B are not isomorphic, then (Est ) = 0 so that we have (2.6). If A = B,

then (Est ) = cst Im (cst are to be determined) so that (Est )ij = cst ij . Then
X
cst ij = ais ( 1 )atj (). (2.7)
G
Set i = j in (2.7) and sum over i to have

n
XX X X
ncst = ati ()ais ( 1 ) = (A()A( 1 ))ts = (In )ts = |G|st .
G i=1 G G
|G|
So cst = n st and substitute it into (2.5) to yield the desired result.
Theorem 2.3.5. Let A() = (aij ()) and B() = (bij ()) be two irreducible
matrix representations of G of degree n and m respectively. Then for any G,
(
X |G|
1
ais ( )btj () = n aij ()st if A = B
G
0 if A and B not isomorphic
Proof. Multiply both sides of (2.5) by ajk () and sum over j:

X n
X n
|G| X |G|
ais ( 1 ) atj ()ajk () = st ij ajk () = aik ()st ,
j=1
n j=1
n
G
P
i.e., G ais ( 1 )atk () = |G|
n aik ()st .
Similarly multiply both sides of (2.6) by bjk () and sum over j to have the
second part of the theorem.
Each entry of A() can be viewed as a function G C. The totality of
A() as runs over G is viewed as a |G|-dimensional complex vector.
Theorem 2.3.6. Let A() = (aij ()) be an irreducible degree n matrix repre-
sentation of G. Then the n2 functions aij : G C are linearly independent.
Proof. Let cij C such that

n
X
cij aij () = 0, for all G.
i,j
Multiply both side by apq ( 1 ) and sum over :

n
X X n
X |G| |G|
0= cij apq ( 1 )aij () = cij pj qi = cpq .
i,j i,j
n n
G
Problems
1. Prove that the irreducible representations of G are of degree 1 if and only

if G is abelian.
2. Let A : G GLn (C) be an irreducible representation of a finite group G.

Prove that n2 |G|.
3. Let A1 () = (a1ij ()), . . . , Ar () = (arij ()) be non-isomorphic represen-

tations of G with degrees
Pr n1 , . . . , nr respectively. Prove thatPthere are
r
linearly independent t=1 n2t functions atij : G C such that t=1 n2t
|G|.
1. Suppose that G is abelian and let A : G GLn (C) be irreducible. By

Corollary 2.3.3, since A()A() = A()A() for all we have A() = c In
for all G; so n = 1 because A is irreducible.
(Daniel): Indeed {A() : G} is a commuting family of matrices, they
are simultaneously diagonalizable. So n = 1.
(Daniel) Conversely suppose that all irreducible representations are of
degree 1. The regular representation of G is injective, i.e., Q(G) = G and
is isomorphic to a diagonal representation T , i.e., T () is diagonal for all
G. Clearly T () for all G are commuting. We then see that
Q(G), and thus G is abelian.
2. By Theorem 2.3.6 and the remark before it.

2.4. CHARACTERS 43
Pr
3. (Roy) The second claim t=1 n2t |G| follows from the first one, because
dim C[G] = |G|. Suppose there are ctij C such that
r X
X
ctij atij () = 0, for all G.
t=1 i,j
Multiply (on the right) both sides by aspq ( 1 ) and sum over , by The-
orem 2.3.5
r X
X X |G| X s s |G| X s s
0= ctij atij ()aspq ( 1 ) = cij aiq ()jp = c a ()
t=1 i,j
ns i,j ns i ip iq
G
for all G. Since the above equation holds for all s = 1, . . . , r and
1 p, q ns , we conclude by Theorem 2.3.6 again that csij = 0 for all
1 i, j ns . Thus all csij s are zero by induction hypothesis.
2.4 Characters
In this section, we assume that G is a finite group.
Let A be a (matrix or linear) representation of G. The function : G C
defined by
() := tr A()
is called a character of G. If A is irreducible, we call an irreducible char-
acter. Characters are class functions on G, i.e., ( ) is a constant for [],
where
[] = { 1 : G}
denotes the conjugacy class in G containing G. So characters are indepen-
dent of the choice of basis in the representation. Though non-similar matrices
could have same trace, we will see in this section that characters completely de-
termine the representation up to isomorphism, i.e., representations having the
same characters if and only if they are isomorphic.
Clearly (e) = tr A(e) = tr In = n where n = (e) is the degree of the
representation. When (e) = 1, is said to be linear; otherwise it is nonlinear.
Example 2.4.1. 1 is a character of G and is called the principal character.

When G = Sm , the sign function is called the alternating character.
We denote by I(G) the set of irreducible characters of G. We will see that

I(G) is a finite set.
Theorem 2.4.2. (Orthogonal relation of the first kind) Let , I(G). Then
(
1 X 1 1 if =
()( ) = (2.8)
|G| 0 if 6=
G
Proof. Let A and B be the irreducible representations of , I(G). Then the

left side of (2.8) is
1 XX
aii ()bjj ( 1 ) (2.9)
|G| i,j
G
If A and B are not isomorphic, from P(2.5), (2.9) is zero. So if 6= , then A

1 1
and B are not isomorphic so that |G| G ()( ) = 0. If = , then we
may assume that A = B. Applying (2.5),
1 X 1 X |G|
()( 1 ) = ij = 1.
|G| |G| i,j n
G
Immediately from Theorem 2.4.2 non-isomorphic irreducible representations

have different characters.
Theorem 2.4.3. Let , I(G). For any G,
(
()
1 X 1 if =
( )() = (e)
|G| 0 if 6=
G
Proof. Similar to the proof of Theorem 2.4.2 and use Theorem 2.3.5.
Theorem 2.4.4. For any character of G, ( 1 ) = () for all G.
Proof. From Problem 2.3, is obtained from a unitary representation, i.e.,
A( 1 ) = A()1 = A() . So
( 1 ) = tr A( 1 ) = tr (A() ) = tr A() = ().
We remark that when is linear, ( 1 ) = ()1 because () = A().

Moreover from Theorem 2.4.4 |()| = 1 for all G. But this is not true for
nonlinear characters.
Let C[G] be the set of all functions f : G C. Notice that C[G] is a vector
space under natural operations and dim C[G] = |G|.
1
P
Theorem 2.4.5. On C[G], (f, g) := |G| G f ()g() is an inner product.
Proof. Exercise.
We sometimes denote it as (, )G . With respect to this inner product, from
Theorem 2.4.2 and Theorem 2.4.4 the irreducible characters of G are orthonor-
mal in C[G] and thus are linearly independent. So we have finitely many irre-
ducible characters of G. Denote them by
I(G) = {1 , . . . , k }.
2.4. CHARACTERS 45
From (2.4.2)
(i , j ) = ij . (2.10)
A function f : G C is said to be a class function if f is constant on each
conjugacy class of G. Denote by C(G) C[G] the set of all class functions. It
is clear that C(G) is a vector space. Since G is a disjoint union of its conjugacy
classes, dim C(G) is the number of conjugacy classes in G.
Theorem 2.4.6. (Isomorphism criterion) Let A, B be two representations of

G. Then A and B are isomorphic if and only if A = B .
Proof. It suffices to establish the sufficiency. By Theorem 2.2.3 we can decom-

pose A and B into their irreducible direct sums:
A() = A1 () Al (), B() = B1 () Bm ()
Among A1 , . . . , Al , B1 , . . . , Bm , denote by D1 , . . . , Dt be the non-isomorphic

ones and let i be the characters of Di , i = 1, . . . , t. Suppose there are ri repre-
sentations among A1 , . . . , Al and si representations among B1 , . . . , Bm isomor-
phic to Di , i 1, . . . , t. Then
Xl t
X t
X
A () = tr ( Ai ()) = ri tr Di () = ri i () (2.11)
i=1 i=1 i=1
Similarly
t
X
B () = si i ().
i=1
From A = B we have
t
X
(ri si )i () = 0, G.
i=1
Since 1 , . . . , t are linearly independent, ri = ti for all i = 1, . . . , t. So A and

B are isomorphic.
Theorem 2.4.7. (Character criterion) C(G) is character of G if and only

Pk
if = i=1 mi i , where m1 , . . . , mk are nonnegative integers.
Proof. From (2.11).
The following theorem implies that the irreducibility of a representation can

be deduced from its character.
Theorem 2.4.8. (Irreducibility criterion) A character of G is irreducible if

and only if (, ) = 1.
Proof. Suppose that A, A1 , . . . , Ak are the representations

Pk for and 1 , . . . , k
respectively and I(G) = {1 , . . . , k }. So A = i=1 m i A i . From Theorem
Pk
2.4.7, = i=1 mi i . Since irreducible characters are orthonormal,
k
X k
X n
X
(, ) = ( mi i , mi i ) = m2i .
i=1 i=1 i=1
Now irreducible, i.e., A irreducible, if and only if only one mi = 1 and other
mi are zeros. In other words, irreducible, if and only if (, ) = 1.
The regular representation Q : G GLn (C), where n = |G|, contains all
the irreducible representations of G as sub-repesentations. We now study its
character structure.
Theorem 2.4.9. The multiplicity of i in Q is i (e), i.e., the degree of i .
k
X
Q = (e)i . (2.12)
i=1
Proof. Recall that if G = {1 , . . . , n }, then the regular representation Q : G

GLn (C) is given by
Q()ij = i ,j .
Let Q be its character. Then
n n
(
X X |G| if = e
Q () = Q()ii = i ,j = (2.13)
i=1 i=1
0 if 6= e
Pk
From Theorem 2.4.7, Q = i=1 mi i , where I(G) = {1 , . . . , k }. Then by
(2.10) and (2.13)
1 X
mi = (Q , i ) = Q ()i () = i (e).
|G|
G
We have the following important relation between the order of G and the
degrees of all irreducible characters.
Pk P
Theorem 2.4.10. |G| = i=1 i (e)2 = I(G) (e)2 .
Pk
Proof. Use (2.13) and (2.12) to have |G| = Q (e) = i=1 i (e)2 .
Immediately we have an upper bound for the degree of any irreducible char-
acter which is not the principal character: , i.e. (e) (|G| 1)1/2 .
Example 2.4.11. |S3 | = 6 so there are three irreducible characters: the prin-
cipal character, the alternating character and an irreducible character of degree
2 (see Example 2.2.1(e)). It is because 6 = 1 + 1 + 22 .
2.4. CHARACTERS 47
Theorem 2.4.12. 1. (Completeness of characters) The irreducible charac-

ters form an orthonormal basis of C(G).
2. The number of irreducible characters |I(G)| of G is the number of conju-
gacy classes of G.
Proof. It suffices to establish the first statement.
Since the space C(G) of class functions on G is the number of conjugacy
classes in G and I(G) = {1 , . . . , k } is a linearly independent set in C(G), it
suffices to show that every class function is a linear combination of 1 , . . . , k .
Let At be the corresponding irreducible representation of t for all t = 1, . . . , k,
and denote by At () = (atij ). Let i (e) = ni for all i. So from Theorem 2.4.10
Pk
|G| = t=1 n2i . From Problem 3.3 the |G| functions atij are linearly independent
and dim C[G] = |G|. So
{atij : i, j = 1, . . . , n, t = 1, . . . , k}
is a basis of C[G]. Let f C(G) C[G]. Then for some ctij , i, j = 1, . . . , n and
t = 1, . . . , k,
k X
X nt
f () = ctij atij (), for all G.
t=1 i,j=1
Since f is a class function,

1 X
f () = f ( 1 ) = f ( 1 )
|G|
G
k nt
1 X X X
= ctij atij ( 1 )
|G|
G t=1 i,j=1
k nt nt
1 XX X X
= ctij atip ( 1 )atpq ()atqj ()
|G|
G t=1 i,j=1 p,q=1
k n
!
X X t
1 X t 1 t
t t
= cij apq () aip ( )aqj ()
t=1 i,j,p,q=1
|G|
G
k
X nt
X 1
= ctij atpq () ij pq by (2.5)
t=1 i,j,p,q=1
nt
nt
k X
X 1
= ctii atpp ()
t=1 i,p=1
nt
k
nt
!
X 1 X
= ct t ()
t=1
nt i=1 ii
a linear combination of 1 , . . . , k .
Example 2.4.13. Since |S4 | = 24 and there are 5 conjugacy classes, there are
5 irreducible characters of G: principal character, alternating character, two
irreducible characters of degree 3 and one irreducible character of degree 2 as
24 = 1 + 1 + 22 + 32 + 32 (the principal and alternating characters are the only
linear characters).
Let [] be the conjugacy class in G containing .
Theorem 2.4.14. (Orthogonal relation of the second kind)
(
|[]| X 1 if [] = []
()() = (2.14)
|G| 0 if [] 6= []
I(G)
Proof. Let I(G) = {1 , . . . , k }. From Theorem 2.4.12, k is also the number

of conjugacy classes. Let 1 , . . . , k G be from the distinct classes. Let
U = (uij ) Ckk be defined by
1/2
|[j ]|
uij = i (j ), i, j = 1, . . . k.
|G|
Then
k
X k
1 X 1 X
(U U )ij = uit ujt = i (t )j (i )|[t ]| = i ()j () = ij ,
t=1
|G| t=1 |G|
G

i.e., U U = I and thus U U = I. So
k
X k
(|[i ]||[j ]|)1/2 ) X
ij = (U U )ij = uti utj = t (i )t (j ) (2.15)
i=1
|G| t=1
and (2.15) is another form of (2.14).

Theorem 2.4.15.
(
1 X 1 if = e
(e)() = (2.16)
|G| 0 if 6= e
I(G)
Proof. Take = e in (2.14).

Theorem 2.4.16. If I(G), then (e) divides |G|.
Proof. Omitted
The above theorem has some interesting implications. For example, if |G|
is a prime, the all characters of G are linear and thus G must be abelian from
Problem 3.1.
Problems
2.4. CHARACTERS 49
1. Let , be characters of G. Prove
(a) () = (), for all , G.

(b) |()| (e), for all G.
(c) (, ) is a nonnegative integer.
2. Let I(G). Prove that
(a) (
X |G| if 1
() =
G
0 otherwise
P
(b) G |()|2 = |G|.
P |G|
3. Prove that I(G) |()|2 = |[]| for any G.
4. Prove that 1 and = are the only linear characters of Sn .
5. What is |I(S3 )|? Find I(S3 ).
6. Let A : G GLn (C) be a representation of G. Show that C() :=

A( 1 )T and D() := A() are representations and they are isomorphic.
Is E() := A() a representation of G?
7. Show that if is a character, so is . Is it true that I(G) if I(G)?
8. Find all the irreducible characters of the alternating group A3 (A3 is iso-
morphic to the cyclic group C3 ).
1. Let I(G) = {1 , . . . , k }. Let , be characters of G with representations

A and B.
(a) () = tr A() = tr A()A() = tr A()A() = tr A() =

().
(b) Since A() can be viewed as a unitary matrix for all G, and
diagonal entries of a unitary matrix have moduli no greater than 1,
we have |()| = |tr Aii ()| tr |Aii ()| (e).
Pk Pk
(c) By Theorem 2.4.12, = i=1 ni i and = i=1 mi i where ni
Pk
and mi are nonnegative integers for all i. So (, ) = i=1 ni mi is
a nonnegative integer.
2. (a) Use Theorem 2.4.2 with 1.

(b) Use Theorem 2.4.2 with = and = e.
3. From Theorem 2.4.14 with = .

4. The transpositions form a conjugacy class [(12)] in Sn (check!). Notice that
if is linear, then it is a representation so that 2 (12) = ((12)2 ) = 1.
Hence ( ) = 1 for all transpositions Sn . Now Sn is generated by
the transpositions and () = ()(), 1 and = are the only
linear characters of Sn .
(Alex) Let be a linear character of Sn , i.e., : Sn C where C = C \
{0} is a homomorphism. Since Im is abelian, the commutator subgroup
[Sn , Sn ] = {xyx1 y 1 : x, y Sn } of Sn is contained in Ker . But the
commutator group [Sn , Sn ] is the alternating group An . Since An is the
largest proper normal subgroup of Sn , Ker is either Sn or An . When
Ker = Sn , 1; when Ker = An , then (12) = 1 since (12)2 = 1
implies that 2 (12) = 1.
5. The conjugacy class of S3 are:
[e] = {e}, [(12)] = {(12), (23), (13)}, [(123)] = {(123), (132)}
So |I(S3 )| = 3 by Theorem 2.4.12. Notice that 1 1 and 2 = are the

linear characters of S3 . Let 3 be the last irreducible character. So from
Theorem 2.4.10, 6 = 21 (e) + 23 (e) + 23 (e) = 1 + 1 + 23 (e) so that 3 is
of degree 2. It is the character corresponding to Problem 2.2.
6. C() = A(()1 )T = (A( 1 )A( 1 ))T = A(()1 )T A(()1 )T =
C()C() for all , G. D() = A() = A()A() = A() A() =
D()D(). By Theorem 2.4.6 C and D are isomorphic since D =
tr D() = tr C() = C . However E() = A() = (A()A()) =
A() A() 6= E()E().
7. From Problem 6: D is a representation. is irreducible if and only if is
irreducible. Just use (, ) = 1, i.e., Theorem 2.4.8.
8. |A3 | = |S3 |/2 = 3. The conjugacy classes of A3 are
[e] = {e}, [(123)] = {(123)}, [(132)] = {(132)}.
Let 1 be the principal character. By Theorem 2.4.10 there are only

two irreducible characters left, both of degree 1, 2 , 3 , say. So 2 (e) =
3 (e) = 1 and |2 ()| = |3 ()| = 1 for all A3 . Since 1 , 2 , 3 are
orthonormal, 2 (123) = , 2 (132) = 2 , 2 (123) = 2 , and 2 (132) = ,
where is the primitive cubic root of unity. Notice that 2 = 3 .
Chapter 3
Multilinear maps and

tensor spaces
3.1 Multilinear maps and tensor maps

Let V1 , . . . , Vm , W be m + 1 vector spaces over a field F. A map : V1
Vm W is called m-multilinear or simply multilinear if
(v1 , . . . , vi + 0 vi0 , . . . , vm ) = (v1 , . . . , vi , . . . , vm ) + 0 (v1 , . . . , vi0 , . . . , vm ),
for all i = 1, . . . , m, i.e., T is linear while restricted on the ith coordinate for all
i. Indeed the condition can be weakened to
(v1 , . . . , vi + vi0 , . . . , vm ) = (v1 , . . . , vi , . . . , vm ) + (v1 , . . . , vi0 , . . . , vm )
for all i. A linear map T : V W can be viewed as a 1-multilinear map. There

is some significiant difference between a linear map and a multilinear map.
Consider the linear map T Hom (V1 V2 , W ) and a multilinear map :
V1 V2 W . Since T is linear,
T (v1 +v10 , v2 +v20 ) = T (v1 , v2 )+T (v10 , v20 ) = T (v1 , 0)+T (0, v2 )+T (v10 , 0)+T (0, v20 ).
Since is multilinear,
(v1 +v10 , v2 +v20 ) = (v1 , v2 +v20 )+(v10 , v2 +v20 ) = (v1 , v2 )+(v1 , v20 )+(v10 , v2 )+(v10 , v20 ).
In particular (v1 , 0) = (0, v2 ) = 0, but T (v1 , 0) and T (0, v2 ) are not necessar-
ily zero.
Example 3.1.1. The following maps are multilinear.
(a) f : C C C defined by f (x, y) = xy.
(b) : V V C defined by (f, v) = f (v).
51
52 CHAPTER 3. MULTILINEAR MAPS AND TENSOR SPACES
(c) : Cm Cn C defined by (x, y) = xT Ay, where A Cmn is given.

(d) : Cm Cn Cmn defined by (x, y) = xy T .
(e) det : Cn Cn C defined by det(x1 , . . . , xn ) = det A, where A =
[x1 xn ].
Qm
(f) f : V1 Vm C defined by f (v1 , . . . , vm ) = i=1 fQ i (vi ), where
m
fQi Vi , i = 1, . . . , Q m, are given. We will write f = i=1 fi . So
m m
i=1 f i (v1 , . . . , vm ) = i=1 fi (vi ).

Qm
(g) g : V1 Vm C defined by g(f1 , . . . , fm ) = i=1 fi (vi ) where
vi V , i = 1, . . . , m, are given.
(h) Let : V1 Vm W and : V1 Vm W . Then +
is also multilinear. So the set of multilinear maps M (V1 , . . . , Vm ; W ) is a
vector space.
Recall the notation in Chapter 1:
:= (n1 , . . . , nm ) := { : = ((1), . . . , (m)), 1 (i) ni , i = 1, . . . , m}
Qm
with || = i=1 ni . We will use frequently. Notice that we can order
according to the lexicographic order. Moreover
Y ni
m X m
XY
aij = ai(i) .
i=1 j=1 i=1
Let Ei = {ei1 , . .P
. , eini } be a basis of Vi , i = 1, . . . , m. So each vi Vi can
ni
be written as vi = j=1 aij eij , i = 1, . . . , m. Let : V1 Vm W be
multilinear. From definition,
n1
X nm
X
(v1 , . . . , vm ) = ( a1j1 e1j1 , . . . , amjm emjm )
j1 =1 jm =1
n1
X nm
X
= a1j1 amjm (e1j1 emjm )
j1 =1 jm =1
X
= a1(1) am(m) (e1(1) , . . . , em(m) )

X
= a (e ) (3.1)

where
m
Y
a := ai(i) C,
i=1
e := (e1(1) , . . . , em(m) ) m
i=1 Vi , (3.2)
for all (n1 , . . . , nm ). (e ) in (3.1) completely determine the multilinear
.
3.1. MULTILINEAR MAPS AND TENSOR MAPS 53
Theorem 3.1.2. (Multilinear extension) Let Ei = {ei1 , . . . , eini } be a basis of

Vi , i = 1, . . . , m. There exists a unique multilinear map : V1 Vm W
such that (e ) = w for all (n1 , . . . , nm ), where e is given in (3.2) and
w W are given.
Proof. Since we want (e ) = w for all , from (3.1) we need to define
X
(v1 , . . . , vm ) = a w

Pni 0
Pni 0
where a , w are given in (3.2) and vi = j=1 aij eij . Let vi = j=1 aij eij
where i = 1, . . . , m. From the definition of ,
(v1 , . . . , vi + cvi0 , . . . , vm )
X
= a1(1) (ai(i) + ca0i(i) ) am(m) w

X X
= a1(1) ai(i) am(m) w + c a1(1) a0i(i) am(m) w

= (v1 , . . . , vi , . . . , vm ) + c(v1 , . . . , vi0 , . . . , vm )

i.e., is multilinear. We are going to show that (e ) = w for all . For
any , write
nj
X
ei(i) = (i)j eij .
j=1
From the definition of and e = (e1(1) , . . . , em(m) ), we have

X X
(e ) = (e1(1) , . . . , em(m) ) = (1)(1) (m)(m) w = w = w .

Thus we just established the existence.

Suppose that there is another multilinear map : V1 Vm W such
that (e ) = w , . Then from (3.1)
X X
(v1 , . . . , vm ) = a (e ) = a w = (v1 , . . . , vm ).

So = .
Let us point some more differences between linear and mutlilinear maps.
When T : V W is linear, T is completely determined by the n := dim V
values T (e1 ), . . . , T (en ) where E =Q{e1 , . . . , en } is a basis of V . But for a
m
multilinear map , wePneed || = i=1 dim Vi values. It is much more than
m
dim(V1 Vm ) = i=1 dim Vi .
Recall Example 3.1.1(d) with m = n = 2, i.e., : C2 C2 C22 defined
by (x, y) = xy T . Since rank (xy T ) min{rank x, rank y T } 1, we have
det((x, y)) = 0. But if x1 := (1, 0)T and x2 := (0, 1)T , then
det((x1 , x1 ) + (x2 , x2 )) = det I2 = 1.
So (x1 , x1 ) + (x2 , x2 ) 6 Im . So Im is not a subspace.

In general the image Im = {(v1 , . . . , vm ) : vi Vi , i = 1, . . . , m} is not
necessarily a subspace of W . The rank of ,
rank = dimhIm i.
Qm
Clearly rankQ i=1 dim Vi . The multilinear map is called a tensor map
m
if rank = i=1 dim Vi . In other words, a tensor map is a multilinear map with
maximal image span. Example 3.1.1 (a) is a trivial tensor map.
Theorem 3.1.3. The multilinear map : V1 Vm P is a tensor map
if and only if the set {(e ) : } is linear independent, where e is given in
(3.2).
Qm
Proof. From (3.1) h(e ) : i = hIm i and || = i=1 dim Vi .
Theorem 3.1.4. Tensor map of V1 , . . . , Vm exists, i.e., there are W and :
V1 Vm W such that is a tensor map.
Qm
Proof. By Theorem 3.1.2, simply pick W so that dim W = i=1 dim Vi and let
{w : } be a basis so that w ( ) determine the multilinear which
is obviously a tensor map.
Clearly tensor maps on V1 Vm are not unique.
The study of multilinear map is reduced to the study of some linear map
(not unique) via a tensor map. A multilinear map : V1 Vm P
is said to have universal factorization property if for any multilinear map
: V1 Vm W , there is T Hom (P, W ) such that = T .

V1 MVm /P
MMM ~~
MMM ~~
MMM ~~
M& ~~~ T
W
Theorem 3.1.5. The multilinear map : V1 Vm P is a tensor map

if and only if has universal factorization property.
Proof. Suppose that is a tensor map. Then {(e ) : } is a basis of
hIm i. From Theorem 1.1.1, there is a unique T1 Hom (hIm i, W ) such that
T1 (e ) = (e ), . Since hIm i P , there is T Hom (P, W ) such that
T |hIm i = T1 . So T (e ) = (e ), . Since T and are multilinear
maps on V1 Vm (Problem 2), from Theorem 3.1.2, T = .
Conversely, suppose that has universal factorization property.
QmIn partic-
ular consider a tensor map on V1 Vm , i.e., dimhIm i = i=1 dim Vi .
Then T = for some linear T . Thus T (hIm i) = hIm i. Hence
dimhIm i dimhIm i.
Qm
So rank = i=1 dim Vi and is a tensor map.
3.2. TENSOR PRODUCTS AND UNIQUE FACTORIZATION 55
Problems
1. Suppose that : W1 Wm W is multilinear and Ti : Vi Wi is
linear, i = 1, . . . , m. Define : V1 Vm W by (v1 , . . . , vm ) =
(T1 v1 , . . . , Tm vm ). Show that is multilinear.
2. Prove that if : V1 Vm W is multilinear and T : V W is
linear, then T is multilinear.
3. Show that when n > 1, the determinant function det : Cn Cn C
is not a tensor map.
4. Suppose that the multilinear map : V1 Vm P has universal
factorization property. Show that the linear map T is unique if and only
if hIm i = P .
1.
(v1 , . . . , vi + cvi0 , . . . , vm )
= (T1 v1 , . . . Ti (vi + cvi0 ), . . . , Tm vm )
= (T1 v1 , . . . , Ti (vi ) + cTi (vi0 ), . . . , Tm vm )
= (T1 v1 , . . . , Ti (vi ), . . . , Tm vm ) + c(T1 v1 , . . . , Ti (vi0 ), . . . , Tm vm )
= (v1 , . . . , vi , . . . , vm ) + c(v1 , . . . , vi0 , . . . , vm ).
2.
T (v1 , . . . , vi + cvi0 , . . . , vm )
= T ((v1 , . . . , vi , . . . , vm ) + c(v1 , . . . , vi0 , . . . , vm ))
= T (v1 , . . . , vi , . . . , vm ) + cT (v1 , . . . , vi0 , . . . , vm )
3. dim(Cn Cn ) = nn but dimhIm deti = dim C = 1. So det : Cn

Cn C is not a tensor map.
4. Suppose that T1 = T2 where T1 , T2 : P W . Then T1 z = T2 z
for all z Im . Since hIm i = P , we have T1 z = T2 z for all z P . So
T1 = T2 .
3.2 Tensor products and unique factorization

Let P be a vector space. If there is a tensor map : V1 Vm P such
that hIm i = P , then P is said to be the tensor product of V1 , , Vm and
is written as m
i=1 Vi = V1 Vm . Clearly
n
Y
dim m
i=1 Vi = dimhIm i = dim Vi .
i=1
The elements of m
i=1 Vi are called tensors. The tensors of the form
(v1 , . . . , vm ) = v1 vm
are called decomposable tensors, i.e., tensor in = are decomposable. The

decomposable tensors span m i=1 Vi so we can find a basis of decomposable
tensors for m V
i=1 i . The tensors which are not decomposable, i.e., tensors in
hIm i \ Im are called indecomposable tensors.
From Theorem 3.1.4:
Qn
Theorem 3.2.1. Let P be a vector space with dim P = i=1 dim Vi . Then
there exists : V1 Vm P so that P is the tensor space.
Theorem 3.2.2. Tensor spaces of V1 , . . . , Vm are isomorphic, i.e., if P = m i=1 Vi
and Q = m
i=1 V i are tensor spaces of V 1 , . . . , Vm , then there exists an invertible
T Hom (P, Q) such that T = .
Proof. Since P and Q are tensor spaces of V1 , . . . , Vm , dim P = dim Q,
{e : }, {e : }
are bases of P and Q. By Theorem 3.1.5 (universal factorization property) there

is T Hom (P, Q) such that T (e ) = e , . Thus T is invertible and
T = .
So all tensor products are isomorphic and thus we use V1 Vm to denote
any one of those.
Theorem 3.2.3. Let : V1 Vm W be a multilinear map. Then there
exists a unique linear T : m
i=1 Vi W such that = T , i.e.,
(v1 , . . . , vm ) = T (v1 , . . . , vm ) = T v .
Proof. From Problem 3.1 #.4.

In other words, the tensor map on m
i=1 Vi has unique factorization prop-
erty.
Corollary 3.2.4. Let : V1 Vm P be a multilinear map. Then is a
tensor map and hIm i = P if and only if has a unique universal factorization
property.
Problems
1. Show that if some vi = 0, then v1 vm = 0.
Pk
2. Let z U V so that z can be represented as z = i=1 ui vi . Prove that
if k is the smallest natural number among all such representations, then
{u1 , . . . , uk } and {v1 , . . . , vk } are linear independent sets respectively.
3.2. TENSOR PRODUCTS AND UNIQUE FACTORIZATION 57
3. Suppose that e1 , e2 V are linearly independent. Prove that e1 e2 +

e2 e1 V V is indecomposable.
4. Let P = mi=1 Vi with tensor map and let T Hom (P, Q) be invertible
linear map. Prove that = T is also a tensor map and Q is the tensor
space with tensor map .
1. The tensor map is a multilinear map. So v1 vm = 0 if vi = 0 for
some i.
2. (Daniel and Roy) Without loss of generality, we suppose on the contrary
Pk1
that {u1 , . . . , uk } is not linearly independent and uk = i=1 ai ui for some
ai C. Then
k
X k1
X k1
X k1
X
z= ui vi = ui vi + (ai ui ) vk = ui (vi + ai vk ),
i=1 i=1 i=1 i=1
which implies that the rank of z is no greater than k 1, a contradiction.

3. Extend e1 , e2 to a basis {e1 , . . . , en } of V . If e1 e2 + e2 e1 were
decomposable, then we would have
n
X Xn X
e1 e2 + e2 e1 = ( i ei ) ( i ei ) = i j ei ej .
i=1 i=1 i,j
Since ei ej , i, j = 1, . . . , n, are linearly independent, 1 2 = 2 1 = 1

and 1 1 = 2 2 = 0 which is impossible.
4. (Roy) Since is multilinear and T is linear, for each i = 1, . . . , m, we have
(v1 , . . . , vi + avi0 , . . . , vm )
= T ((v1 , . . . , vi , . . . , vm ) + a(v1 , . . . , vi0 , . . . , vm ))
= T ((v1 , . . . , vi , . . . , vm )) + aT ((v1 , . . . , vi0 , . . . , vm ))
= (v1 , . . . , vi , . . . , vm ) + a(v1 , . . . , vi0 , . . . , vm )
Qm
Thus is multilinear. Notice that rank = rank = i=1 dim Vi , since
T is invertible. Therefore, is a tensor map and Q is the corresponding
tensor space.
(Daniel) By Theorem 3.1.5, in order to show that is a tensor map,
we need to show that has universal factorization property. Let f :
m m
i=1 Vi R be any multilinear map. Since : i=1 Vi P is a tensor
map and P = i=1 Vi , there is f : P R such that f = f . Set
m
f = f T 1 : Q R. Then
f = f = f T 1 T = f .
3.3 Basic properties of tensors and induced in-

ner products
Theorem 3.3.1. If u1 um + + v1 vm = 0, then (u1 , . . . , um ) +
+ (v1 , . . . , vm ) = 0 for any multilinear : V1 Vm W .
Proof. From Theorem 3.2.3, = T for some T Hom (m
i=1 Vi , W ). So
(u1 , , um ) = T (u1 , , um ) = T (u1 um )
and similarly (v1 , . . . , vm ) = T (v1 vm ). Hence
(u1 , . . . , um ) + + (v1 , . . . , vm )
= T (u1 um ) + + T (v1 vm )
= T (u1 um + + v1 vm )
= 0.
Theorem 3.3.2. Let vi Vi , i = 1, . . . , m. Then v1 vm = 0 if and only

if some vi = 0.
Proof. One implication is trivial. Suppose v1 vm = 0 but vi 6= 0 for all
i. Then there were fi V such that fi (vi ) = 1 for all i. Let
m
Y
= fi : V1 Vm C
i=1
which is multilinear. Then

m
Y
(v1 , . . . , vm ) = fi (vi ) = 1.
i=1
By Theorem 3.3.1 (v1 , . . . , vm ) = 0, a contradiction.

Theorem 3.3.3. u1 Q um = v1 vm 6= 0 if and only if vi = ci ui 6= 0
m
for all i = 1, . . . , m and i=1 ci = 1.
Proof. One implication is trivial. Suppose u1 um = v1 vm 6= 0. From
Theorem 3.3.2 all ui , vi are nonzero. From Theorem 3.3.1 for all multilinear ,
(u1 , . . . , um ) = (v1 , . . . , vm ).
Suppose that uk and vk are not linearly dependent for some k. Then there is
Vk V such that fk (vk ) = 1 Q
and fk (uk ) = 0. For other i 6= k, choose fi Vi
m
such that f (vi ) = 1. Set := i=1 fi . Then
m
Y
(v1 , . . . , vm ) = fi (vi ) = 1
i=1
3.3. BASIC PROPERTIES OF TENSORS AND INDUCED INNER PRODUCTS59
and
m
Y
(u1 , . . . , um ) = fi (ui ) = 0
i=1
contradicting (u1 , . . . , um ) = (v1 , . . . , vm ). Hence vk = cuk for all k. From

m
Y
0 6= u1 um = v1 vm = ( ci )u1 um
i=1
Qm
we have i=1 ci = 1.
Since the decomposable elements span the tensor space m i=1 Vi , each z
m
i=1 Vi is a linear combination of decomposable tensors. Let k be the smallest
number of decomposable tensors in all such linear combinations for z. We call
k the smallest length or the rank of z. The smallest length of z 6= 0 is one if
and only if z is decomposable.
Pk
Theorem 3.3.4. Let z U V so that z can be represented as z = i=1 ui vi .
Then k the smallest length of z if and only if {u1 , . . . , uk } and {v1 , . . . , vk } are
linear independent sets, respectively.
Proof. One implication followsPfrom Problem 3.2 #2.
r
For the sufficiency, let z = j=1 xj yj and we are going to show that k r.
Since v1 , . . . , vk are linearly independent, there is g V such that g(vl ) = 1
(l is arbitrary fixed), g(vj ) = 0 when j 6= l. Let f U be arbitrary. Then
f g : U V C is bilinear. Since
k
X r
X
ui vi = xj yj
i=1 j=1
from Theorem 3.3.1 with = f g, we have

k
X r
X Xr
f (ul ) = f (ui )g(vi ) = f (xj )g(yj ) = f ( g(yj )xj ).
i=1 j=1 j=1
Pr
Since f V is arbitrary, ul = j=1 g(yj )xj , i.e., ul hx1 , . . . , xr i for all
l = 1, . . . , k. Since u1 , . . . , uk are linearly independent, we have k r.
We now consider an induced inner product of m i=1 Vi . Suppose that (, )i is
an inner product on Vi and Ei = {ei1 , . . . , eini } is an orthonormal basis of Vi
for all i = 1, . . . , m. We know that
E := {e
:= e1(1) em(m) : }
is a basis of mi=1 Vi , where = (n1 , . . . , nm ). We would like to have an inner

product (, ) on m i=1 Vi such that E is an orthonormal basis, i.e.,

(e
, e ) = , .
Such inner product is unique from Theorem 1.6.3. From Problem 1.6 #1
X
(u, v) := a b (3.3)

P
P
where u = a e , v
= b e m
i=1 Vi . It appears depending on the
choice of basis but it does not.
Theorem 3.3.5. Let Vi be inner product space with orthonormal basis Ei =
{ei1 , . . . , eini }, i = 1, . . . , m. The inner product obtained from (3.3) satisfies
m
Y
(u1 um , v1 vm ) = (ui , vi )i (3.4)
i=1
ui , vi Vi for all i.
Proof. Since ui , vi Vi we have
ni
X ni
X
ui = aij eij , vi = bij eij , i = 1, . . . , m.
j=1 j=1
Since is multilinear, from (3.1)

m
XY
u = (u1 , . . . , um ) = ai(i) e

i=1
m
XY
v = (v1 , . . . , vm ) = bi(i) e
.
i=1
Qm Pni P Qm
From (3.3) and i=1 j=1 cij = i=1 ci(i) ,
m
XY
(u , v ) = ai(i) bi(i)
i=1
Ym Xni
= aij bij
i=1 j=1
Ym X ni ni
X
= ( aij eij , bij eij )i
i=1 j=1 j=1
Ym
= (ui , vi )i .
i=1
There are many bilinear maps from (ki=1 Vi ) (m m

i=k+1 Vi ) to i=1 Vi since
m
Y k
Y m
Y
dim m
i=1 Vi = ni = ni ni = dim(ki=1 Vi ) dim(m
i=k+1 Vi ). (3.5)
i=1 i=1 i=k+1
What we like is one that maps (v1 vk , vk+1 vm ) to v1 vm .

3.3. BASIC PROPERTIES OF TENSORS AND INDUCED INNER PRODUCTS61
Theorem 3.3.6. There is a unique multilinear map : (ki=1 Vi )(m

i=k+1 Vi )
m
i=1 V i such that
(v1 vk , vk+1 vm ) = v1 vm , (3.6)
and
(V1 Vk ) (Vk+1 Vm ) = V1 Vm . (3.7)
Proof. By (3.5) the tensor map that satisfies (3.6) exists and is unique. From
hIm i = hv1 vm : vi Vi i = m
i=1 Vi , (3.7) follows. See Problem 3.2 #4
for details.
We also write for in Theorem 3.3.6. So (3.7) can be written as
(V1 Vk ) (Vk+1 Vm ) = V1 Vm
and (3.6) can be written as
(v1 vk ) (vk+1 vm ) = v1 vm .
Problems
1. Suppose that
Pkv1 , . . . , vk V are linearly independent and u1 , . . . , uk U .
Prove that i=1 ui vi = 0 if and only if u1 = = uk = 0.
2. Let v1 , . . . , vk V and A Ckk . Suppose AAT = Ik and uj =
Pk Pk Pk
i=1 aij vi , j = 1, . . . , k. Prove that i=1 ui ui = i=1 vi vi .
3. Define : Ck Cn Ckn by x y = xy T . Let Ck and Cn be equipped

with the standard inner products. Prove that for any A, B Ckn =
Ck Cn , the induced inner product is given by (A, B) = tr B A.
4. Let Ei = {ei1 , . . . , eini } be a basis of Vi , i = 1, . . . , m. Define
: (V1 Vk ) (Vk+1 Vm ) V1 Vm
by (e1i1 ekik , ek+1,ik+1 emim ) = e1i emim (with

bilinear extension). Show that is the tensor map satisfying
(v1 vk , vk+1 vm ) = v1 vm .
Pk
5. Let z = i=1 ui vi wi U V W . Prove that if {u1 , . . . , uk } and
{v1 , . . . , vk } are linearly independent and wi 6= 0 for all k, then k is the
smallest length of z.
Solutions to Problem 3.3

Pk Pk
1. By Theorem 3.3.1 i=1 ui vi = 0 implies that i=1 (ui , vi ) = 0 for
any multilinear map . Since v1 , . . . , vk are linearly independent, we can
choose f V such that f (vi ) = 1 for all i. If some ui 6= 0, then there
would be g U such that g(ui ) = 1 and g(uj ) = 0 for j 6= i. The map
Pk
f g : U V C is multilinear so that i=1 f g(ui , vi ) = 1, a contradiction.
2.
k k
k ! k !!
X X X X
uj uj = aij vi aij vi
j=1 j=1 i=1 i=1
k
k k !
X XX
= aij a`j vi v`
j=1 `=1 i=1

k X
X k Xk
= aij a`j vi v`
`=1 i=1 j=1
k X
X k k
X
= i` vi v` = vj vj
`=1 i=1 j=1
3. (Roy) Suppose A = x y and B = z w. Then by Theorem 3.3.5, we

have k ! n
X X
(A, B) = (x, z) (y, w) = xi zi yj wj .
i=1 j=1
On the other hand,

n
k X n
k X
k ! n
X X X X
tr B A = aij bij = xi yj zi wj = xi zi yj wj .
i=1 j=1 i=1 j=1 i=1 j=1
Thus the induced inner product is given by (A, B) = tr B A.

4.
5.
3.4 Induced maps

In this section we study Hom (m m
i=1 Vi , i=1 Wi ). Let Ti Hom (Vi , Wi ). Define
m m
the multilinear map from i=1 Vi to i=1 Wi by
(v1 , . . . , vm ) := T1 v1 Tm vm .
By Theorem 3.2.3, there is a unique T Hom (m m

i=1 Vi , i=1 Wi ) such that
T (v1 vm ) = T1 v1 Tm vm .
3.4. INDUCED MAPS 63
Denote such T by T1 Tm and call it the induced map of T1 , . . . , Tm ,

i.e.,
(m
i=1 Ti )(v1 vm ) = T1 v1 Tm vm .
We will see that in Section 3.7 that T1 Tm is a tensor of m

i=1 Hom (Vi , Wi ).
The main focus in this section is to study T1 Tm Hom (m m
i=1 Vi , i=1 Wi )
as a linear map.
Theorem 3.4.1. Let Si Hom (Wi , Ui ), Ti Hom (Vi , Wi ), i = 1, . . . , m.
Then
(m m m
i=1 Si )(i=1 Ti ) = i=1 (Si Ti ).
Proof.
(m m
i=1 Si )(i=1 Ti )(v1 vm )
= (mi=1 Si )(T1 v1 Tm vm )
= Si Ti v1 Sm Tm vm
= mi=1 (Si Ti )(v1 vm ).
Since mi=1 Vi is spanned by the decomposable elements, we have the desired

result.
Theorem 3.4.2. Let Ti Hom (Vi , Wi ), i = 1, . . . , m. Then
m
Y
rank m
i=1 Ti = rank Ti .
i=1
Proof. Let rank Ti = ki for all i. So there is a basis {ei1 , . . . , eiki , eiki +1 , . . . , eini }
for Vi such that T ei1 , . . . , T eiki are linearly independent in Wi and T eiki +1 =
T eini = 0, i = 1, . . . , m. Notice that {e : (n1 , . . . , nm )} is a basis for
m i=1 Vi . Moreover
(m
i=1 Ti )e = T e1(1) Tm em(m)
so that if 6 (k1 , . . . , km ), then (m

i=1 Ti )e = 0 (because some (i) > ki ).
Since T ei1 , . . . , T eiki are linearly independent in Wi for all i, the vectors
(m
i=1 Ti )e , (k1 , . . . , km )
are linearly independent in m

i=1 Wi (Why?) Hence
m
Y m
Y
rank m
i=1 Ti = |(k1 , . . . , km )| = ki = rank Ti .
i=1 i=1
Theorem 3.4.3. Let Ti Hom (Vi , Wi ), where Vi , Wi are inner product spaces,
i = 1, . . . , m. Then
(m m
i=1 Ti ) = i=1 Ti .
Proof. We use the same notation (, ) for all the inner products on Vi , Wi .
(m
i=1 Ti v , w ) = (T1 v1 Tm vm , w1 wm )
Ym
= (Ti vi , wi )
i=1
Ym
= (vi , Ti wi )
i=1
= (v1 vm , T1 w1 Tm

wm )
m
= (v , i=1 Ti w )
Since m
i=1 Vi is spanned by decomposable tensors, we have the desired result.
Problems
1. Prove that
(a) T1 Tm = 0 if and only if some Ti = 0,

(b) T1 Tm is invertible if and only if Ti are all invertible.
2. Let Si , Ti Hom (Vi , Wi ), i = 1, . . . , m. ProveQthat m m

i=1 Ti = i=1 Si 6= 0
m
if and only if Ti = ci Si 6= 0, i = 1, . . . , m and i=1 ci = 1.
3. Let Ti End Vi , i = 1, . . . , m. Prove that be m i=1 Ti is invertible if

and only if Ti (i = 1, . . . , m) are invertible. In this case (m i=1 Ti )
1
=
m 1
i=1 Ti .
4. Let Ti Hom (Vi , Wi ), i = 1, . . . , m. Define
: Hom (Vi , Wi ) Hom (Vm , Wm ) Hom (m m

i=1 Vi , i=1 Wi )
by (T1 , . . . , Tm ) = T1 Tm . Prove that is multilinear.

m m
1. (a)
Qm Directly from Theorem 3.4.2: i=1 Ti = 0 0 = rank i=1 Ti =
i=1 rank T i , i.e., some T i = 0. The converse is trivial. (b) From Theorem
3.4.2.
Another approach: use Theorem 3.7.2 and Theorem 3.3.2.
3.5. MATRIX REPRESENTATIONS AND KRONECKER PRODUCT 65
Qm
2. (Roy) If Ti = ci Si with i=1 ci = 1, then clearly
T1 Tm (v1 vm ) = T1 v1 Tm vm
= c1 S1 v1 cm Sm vm
Ym
= ( ci )S1 v1 Sm vm
i=1
= S1 Sm (v1 vm )
So m m m m
i=1 Ti = i=1 Si . Conversely if i=1 Ti = i=1 Si 6= 0, then
T1 v1 Tm vm = S1 v1 Sm vm 6= 0
for all v1 vm 6 Ker m i=1 Ti . In particular, if vi 6 Ker Ti for all

i, then by Theorem 3.3.2, v1 vm 6 Q Ker m i=1 Ti . By Theorem
m
3.3.3 Ti vi = ci Si vi for all vi 6 Ker Ti Qwith i=1 ci = 1. By symmetry
m
Si vi = c0i Ti vi for all vi 6 Ker Si with i=1 c0i = 1. So Ti = ci Si in the
complement of Clearly Si vi = c1i Ti vi So Ti and ci Si
Another approach: apply Theorem 3.7.2 and Theorem 3.3.3.
3. The first part follows from Theorem 3.4.2. By Theorem 3.4.1
1 1
(m m m
i=1 Ti )(i=1 Ti ) = i=1 (Ti Ti ) = I.
1
So (m
i=1 Ti )
1
= (m
i=1 Ti ).
4.
(T1 , . . . , Ti + cTi0 , . . . , Tm )(v1 vm )

= T1 Ti + cTi0 Tm (v1 vm )
= T1 v1 (Ti + cTi0 )vi Tm vm
= T1 v1 (Ti vi + cTi0 vi ) Tm vm
= T1 v1 Tm vm + cT1 v1 Ti0 vi Tm vm
= ((T1 , . . . , Ti , . . . , Tm ) + c(T1 , . . . , Ti0 , . . . , Tm ))(v1 vm )
So (T1 , . . . , Ti +cTi0 , . . . , Tm ) = ((T1 , . . . , Ti , . . . , Tm )+c(T1 , . . . , Ti0 , . . . , Tm )).
3.5 Matrix representations of induced maps and

Kronecker product
Let Ei = {ei1 , . . . , eini } be a basis of Vi and Fi = {fi1 , . . . , fiki } be a basis of
Wi , i = 1, . . . , m. Then
E = {e
: (n1 , . . . , nm )}, F = {f : (k1 , . . . , km )}
are bases for m m

i=1 Vi and i=1 Wi respectively, where (n1 , . . . , nm ) and (k1 , . . . , km )
are in lexicographic ordering.
Theorem 3.5.1. Let Ti Hom (Vi , Wi ) such that Ai = [Ti ]F i

Ei = (ast ) Cki ni
i
i = 1, . . . , m. That is
ki
X
Ti eit = aist fis , t = 1, . . . , ni , i = 1, . . . , m.
s=1
Then the matrix representation of m

i=1 Ti with respect to E and F is given
by m !
F
Y
m i
[i=1 Ti ]E = a(i)(i)
i=1 (k1 ,...,km )
(n1 ,...,nm )
F Qm
In other words, the (, ) entry of [m
i=1 Ti ]E is i=1 ai(i)(i) .
Proof. For any (n1 , . . . , nm ), we have

m
i=1 Ti e = T1 e1(1) Tm em(m)
k1
X km
X
= a1s(1) f1s am
s(m) fms
s=1 s=1
X
= a1(1)(1) f1(1) am
(m)(m) fm(m)
(k1 ,...,km )
m !
X Y
= ai(i)(i) f .
(k1 ,...,km ) i=1
Let Ai = (aist ) Cki ni , i = 1, . . . , m. The Kronecker product of

A1 , . . . , Am , denoted by A1 Am C(Qm
i=1 ki )(
Qm
i=1 ni )
, is defined by
m
Y
(m
i=1 Ai ), = ai(i)(i) ,
i=1
where (k1 , . . . , km ), (n1 , . . . , nm ), both arranged in lexicographic

ordering respectively. We immediately have the following
Theorem 3.5.2. Let Ai = [Ti ]F
Ei i = 1, . . . , m. Then
i
F Fi
[m m m
i=1 Ti ]E = i=1 [Ti ]Ei = i=1 Ai .
Example 3.5.3. Let A Cmn and B Cpq . Then

a11 B . . . a1n B
A B = = (aij B) Cmpnq
am1 B amn B
In general A B 6= B A.
3.5. MATRIX REPRESENTATIONS AND KRONECKER PRODUCT 67
Notice that
(m, p) = {(1, 1), . . . , (1, p), (2, 1), . . . , (2, p), , (m, 1), . . . , (m, p)}
and
(n, q) = {(1, 1), . . . , (1, q), (2, 1), . . . , (2, q), , (n, 1), . . . , (n, q)}.
Consider = (1, i) (m, p), = (1, j) (n, q). Then
(A B)(1,i),(1,j) = a(1)(1) b(2)(2) = a11 bij .
That is, the (A B)[1, . . . , p|1, . . . , q] = a11 B. Similar for other blocks.
Problems
1. Let Ai , Bi Cki ,ni , i = 1, . . . , m. Prove the following:
(i) m
i=1 Ai = 0 if and only if some Ai = 0.
m
(ii) Q m
i=1 Ai = i=1 Bi 6= 0 if and only if Ai = ci Bi 6= 0 for all i and
m
i=1 ci = 1.
Qm
(iii) rank (m i=1 Ai ) = i=1 rank Ai .
(iv) (m m
i=1 Ai ) = i=1 Ai .
2. From the definition of Kronecker product prove that (m m

i=1 Ai )(i=1 Bi ) =
m
i=1 Ai Bi . (Hint:
X m
Y Ym m X
Y ti
( a(i)(i) )( bi(i)(i) )
i
= ( ai(i)j bij(i) )
(t1 ,...,tm ) i=1 i=1 i=1 j=1
where Ak = (akij )).

3. Show that if Ai Cni ni , i = 1, . . . , m, are triangular (diagonal, resp.)
matrices, then A1 Am is also triangular
Qm (diagonal, resp.). In par-
ticular the (, )-entry of A1 Am is i=1 ai(i)(i) .
4. Show that if Ai , Bi Cni ni are similar for all i = 1, . . . , m, then m
i=1 Ai
and m B
i=1 i are similar.
5. Which one of x y = xy T and x y = yxT , x Cm and y Cn is the
Kronecker product? Explain.
6. (Merris) Show that for any A, B Cnn , (1) If AB
P= BA, then eA+B =
A B A B AI+B A k
e e and (2) e e = e , where e := k=1 A /k.

1. (i) From Problem 3.4 #1 and Theorem 3.5.2

(Roy) If Ai = 0, then m m
i=1 Ai = 0 because every entry of i=1 Ai has
a factor of some entry of Ai .
(). Suppose on the contrary that Ai 6= 0 for all i. Then each
Ai contains some nonzero entry aisi ti . Let (k1 , . . . , km ) and
(n1 , . . . , nm ) be such that (i) = si and (i) = ti for all i.
It follows that the (, ) entry of m i=1 Ai is nonzero, which means
mi=1 Ai 6= 0.
(ii) From Problem 3.4 #2 and Theorem 3.5.2
(iii) From Theorem 3.4.2 and Theorem 3.5.2
(iv) From Theorem 3.4.3 and Theorem 3.5.2
2. Let A = (aist ) and B = (bikl ). The (, )-entry of (m m
i=1 Ai )(i=1 Bi ) is
X m
Y m
Y
( ai(i)(i) )( bi(i)(i) )
(t1 ,...,tm ) i=1 i=1
X m
Y
= (ai(i)(i) )bi(i)(i) )
(t1 ,...,tm ) i=1
m X
Y ti
= ( ai(i)j bij(i) )
i=1 j=1
which is the (, )-entry of m

i=1 Ai Bi .
3. Recall that
m
Y
(m
i=1 Ai ), = ai(i)(i) ,
i=1
where (k1 , . . . , km ), (n1 , . . . , nm ), both arranged in lexico-

graphic ordering respectively. When < in lexicographic ordering,
(i) < (i) for some i (remark: converse is not true) so that ai(i)(i) = 0,
i.e., (m m
i=1 Ai ), = 0. In other words, (i=1 Ai ) is upper triangular.
4. Suppose Bi = Pi Ai Pi1 for all i. Then from Problem 2

1 1
m
i=1 Bi = m m m m
i=1 (Pi Ai Pi ) = (i=1 Pi )(i=1 Ai )(i=1 Pi )
= (m m m
i=1 Pi )(i=1 Ai )(i=1 Pi )
1
.
5. (Roy) None of them is the Kronecker product since x y Cmn1 .

6. (Roy) Let E = {e1 , e2 , . . . , en } be a basis of Cn so that [A]E
E is upper trian-
gular. Since [I]E
E is also upper triangular, both AI n and In A are upper
triangular with respect to the basis E E in the lexicographic order by
Theorem 3.6.1(a). Therefore, we may assume, without loss of generality,
that A is upper triangular with diagonal entries 1 , . . . , n in order. The
3.6. INDUCED OPERATORS 69
diagonal entries of AIn are 1 , . . . , 1 , 2 , . . . , 2 , . . . , n , . . . , n in order,

and the diagonal entries of In A are 1 , . . . , n , 1 , . . . , n , . . . , 1 , . . . , n
in order. Thus the eigenvalues of AIn In A are i j , i, j = 1, . . . , n.
3.6 Induced operators

We now study some basic properties of the induced map T1 Tm .
Theorem 3.6.1. Let Ti End Vi and let Ai = [Ti ]E
Ei , where Ei is a basis of Vi ,
i
i = 1, . . . , m.
(a) Suppose that Ai is upper triangular with diag Ai = (i1 , . . . , ini ). Then
F
[m m
i=1 Ti ]E = i=1 Ai is also upper triangular with (, )-diagonal entry
Qm
i=1 i(i) , .
(b) If ij , j = Q
1, . . . , ni , are the eigenvalues of Ti , then the eigenvalues of
m
m T
i=1 i are i=1 i(i) , .
(c) If sij , j = 1, . .Q
. , ni , are the singular values of Ti , then the singular values
m
of m T
i=1 i are i=1 si(i) , .
Q m
(d) tr m i=1 Ti = i=1 tr Ti .
Q m Qm
(e) det m i=1 Ti = i=1 (det Ti )
n/ni
, where n := i=1 ni .
Proof. (a) Use Problem 3.4 #3 for matrix approach.
The following is linear operator approach. Since Ai is upper triangular, For
any (n1 , . . . , nm ), we have
Ti ei(i) = i(i) ei(i) + vi ,
where vi heij : j < (i)i. Then
(m
i=1 Ti )e = T1 e1(1) Tm em(m)
= (1(1) e1(1) + v1 ) (m(m) em(m) + vm )
Ym
= i(i) e
+w
i=1
where w is a linear combination of u1 um . Each ui is either ei(i) or vi

and at least one is vi . But vi is a linear combination of eij , j < (i). So u and
thus w is a linear combination of e such that is before in lexicographic
F
order. So [m T ]
i=1 i E is upper triangular.
(b) By Schurs triangularization theorem we may assume that the matrix
representation of Ai is upper triangular for all i (cf. Theorem 3.4.1 and Problem
3.4 #3). Then use (a).
(c) The singular values of A are the eigenvalue nonnegative square roots of
A A. Apply (b) on A A.
P Qm Qm Pni Qm
(d) tr (m
i=1 Ti ) = i=1 i(i) = i=1 j=1 ij = i=1 tr Ti .
(e) Since determinant is equal to the product of eigenvalues,
m
YY m Y
Y
det m
i=1 Ti = i(i) = i(i)
i=1 i=1
Since (i) is chosen from 1, . . . , ni and such (i) appears in ||/ni times in those
n := || elements in . So
Y ni
Y ni
Y
||/ni n/ni
i(i) = ij = ij
j=1 j=1
and hence
n/ni
m
Y ni
Y m
Y
det(m
i=1 Ti ) =
ij = (det Ti )n/ni .
i=1 j=1 i=1
Theorem 3.6.2. Let Ti End Vi where Vi are inner product spaces, i =

1, . . . , m. If all Ti are (a) normal, (b) Hermitian, (c) psd, (d) pd, and (e)
unitary respectively, so is m i=1 Ti .
Proof. Recall from Theorem 3.4.1 (m m m

i=1 Si )(i=1 Ti ) = i=1 (Si Ti ) and Theorem
m m
3.4.3 i=1 Ti = (i=1 Ti ) .
(a) If all Ti are normal, then
(m m m m m m m m
i=1 Ti ) (i=1 Ti ) = (i=1 Ti )(i=1 Ti ) = i=1 Ti Ti = i=1 Ti Ti = (i=1 Ti )(i=1 Ti ) .
So m
i=1 Ti is normal.
(b) If all Ti are Hermitian, then
(m m m
i=1 Ti ) = i=1 Ti = i=1 Ti
i.e., mi=1 Ti is Hermitian.

(c) and (d) If all Ti are psd (pd), then m
i=1 Ti is Hermitian and by Theorem
3.6.1 all eigenvalues of m T
i=1 i are nonnegative (positive). So m
i=1 Ti is psd (pd).
(e) If all Ti are unitary, then
(m m m m
i=1 Ti ) (i=1 Ti ) = i=1 Ti Ti = i=1 IVi = Im
i=1 Vi
.
We now study the converse of Theorem 3.6.2.
Theorem 3.6.3. Let Ti End Vi where Vi are inner product spaces, i =

1, . . . , m.
3.6. INDUCED OPERATORS 71
(a) If m
i=1 Ti 6= 0 is normal, then all Ti 6= 0 are normal.
(b) If
Qm mi=1 Ti 6= 0 is Hermitian, then Ti = ci Hi where Hi are Hermitan and
i=1 ci R.
Proof. (a) From Problem 3.4 #1, Ti 6= 0. There is an orthonormal basis Ei =

{ei1 , . . . , eini } such that [Ti ]E
Ei is upper triangular. Let E = {e : }
i
E
which is an orthonormal basis for m m
i=1 Vi . From Theorem 3.6.1, [i=1 Ti ]E is
E
upper triangular. On the other hand [m
i=1 Ti ]E is a normal matrix. By Lemma
E
1.5.3 [m
i=1 Ti ]E must be diagonal. now
E E1 Em
[m
i=1 Ti ]E = [T1 ]E1 [Tm ]Em 6= 0.
From Problem 3.5 #3, each [Ti ]E Ei is diagonal. Hence Ti is normal for all i.
i
(b) From (a) each Ti 6= 0 and is normal. So some eigenvalue of Ti is nonzero,

i.e., there exists some such that the eigenvalues of Ti are i(i) 6= 0 for
all i. Consider any fixed Tk . Since the eigenvalues of the Hermitian m i=1 Ti are
real, from Theorem 3.6.1(b)
m
Y
( i(i) )kj R, j = 1, . . . , nk
i=1,i6=k
since they are (not all) eigenvalues of m

i=1 Ti 6= 0 and are nonzero. Let
m
Y
ak := i(i) 6= 0
i6=k
since k(k) 6= 0. Then Hk := ak Tk is normal with real eigenvalues, i.e., Hk is

Hermitian. Set ck := a1
k . Then
m
Y
0 6= m
i=1 Ti = ( ck )(m
k=1 Hk ).
k=1
Qm
Since m m
i=1 Ti and k=1 Hk are Hermitian, we have k=1 ck R.
Problems
1. Prove
Qthat if m
i=1 Ti 6= 0 is pd (psd), then Ti = ci Hi and Hi is pd (psd)
m
and i=1 ci > 0.
2. ProveQthat mi=1 Ti 6= 0 is unitary if and only if Ti = ci Ui and Ui is unitary

m
and | i=1 ci | = 1.
3. Suppose Si , Ti End Vi , i = 1, . . . , m, are psd. Show that m

i=1 (Si + Ti )
m
i=1 S i + m
i=1 T i .
4. Show that if the eigenvalues of A Cnn are 1 , . . . , n , then the eigen-

values of A In In A are i j , i, j = 1, . . . , n.
5. Let Ti End V and vi V , i = 1, . . . , m where m dim V . Prove that

(m
i=1 Ti v , v ) = 0 for all (vi , vj ) = ij if and only if some Ti = 0.
1. By Theorem 3.6.2 and Theorem 3.6.3(b).
2. Suppose that mi=1 Ti 6= 0 is unitary. By Theorem 3.6.2, all Ti are normal

and by Problem 3.4 #3, all Ti are invertible. The converse is easy by
Theorem 3.4.1 and Theorem 3.4.3.
3.
4. By Schurs triangularization theorem, there is unitary matrix U such that

U AU = T is upper triangular with diag T = (1 , . . . , n ). The matrix
U I is unitary by Problem 2. So
(U I)(A In In A)(U I)
= (U I) (A In In A)(U I)
= T I I T
whose spectrum is the spectrum of A In In A counting multiplicities.

From Theorem 3.6.1(a), the eigenvalues of T I I T are i j for
all i, j. In particular, A In In A is not invertible.
5.
3.7 Tensor product of linear maps

Given Ti Hom (Vi , Wi ), i = 1, . . . , m, we studied basic properties of the in-
duced map m m m m m
i=1 Ti Hom (i=1 Vi , i=1 Wi ). Not every member of Hom (i=1 Vi , i=1 Wi )
m
is of the form i=1 Ti (see Problem 1), but the span of all the induced maps is
Hom (m m
i=1 Vi , i=1 Wi ).
Theorem 3.7.1. Hom (m m m

i=1 Vi , i=1 Wi ) = hi=1 Ti : Ti Hom (Vi , Wi )i.
Proof. Let Ei = {ei1 , . . . , eini } be a basis of Vi and Fi = {fi1 , . . . , fiki } be a

basis of Wi , i = 1, . . . , m. Then
E = {e
: (n1 , . . . , nm )}, F = {f : (k1 , . . . , km )}
3.7. TENSOR PRODUCT OF LINEAR MAPS 73
are bases for m m

i=1 Vi and i=1 Wi respectively. Set
1 := (n1 , . . . , nm ), 2 := (k1 , . . . , km )
and both are in lexicographic ordering. Let S Hom (m m

i=1 Vi , i=1 Wi ) be
arbitrary. For any 1 ,
X
Se = c f .
2
i
Define Tjq Hom (Vi , Wi ) such that
i
Tjq eip = pq fij ,
where p, q = 1, . . . , ni , j = 1, . . . ki , i = 1, . . . , m. For any 2 , 1 , set

1 m
T := T(1)(1) T(m)(m) .
Then
1 m
T e = T(1)(1) e1(1) T(m)(m) em(m)
= (1)(1) f1(1) (m)(m) fm(m)
= f .
So

X X X X X

c T e = c f = c f = Se
.
2 1 2 1 2
Since {e m
: 1 } is a basis of i=1 Vi we have
X X

S= c T .
2 1

In other words, S is a linear combination of induced maps T ( 2 ,
1 ).
We now see that T1 Tm can be viewed as a (decomposable) tensor

when each Ti Hom (Vi , Wi ) is viewed as a vector, i.e., we have the following
result.
Theorem 3.7.2. There exist a unique tensor map such that Hom (m m
i=1 Vi , i=1 Wi )
is the tensor space of Hom (V1 , W1 ), . . . , Hom (Vm , Wm ), i.e.,
Hom (m m
i=1 Vi , i=1 Wi ) = Hom (V1 , W1 ) Hom (Vm , Wm ).
In other words, T1 Tm = (T1 , . . . , Tm ).

Proof. From Problem 3.4 #4, the map (T1 , . . . , Tm ) = T1 Tm

Hom (m m
i=1 Vi , i=1 Wi ) is multilinear and is unique.
From Theorem 3.7.1, Hom (m m
i=1 Vi , i=1 Wi ) is spanned by the induced maps
m
i=1 Ti , Ti Hom (Vi , Wi ), i.e.,
hIm i = hm m m
i=1 Ti : Ti Hom (Vi , Wi )i = Hom (i=1 Vi , i=1 Wi ).
Moreover
dimhIm i = dim Hom (m m

i=1 Vi , i=1 Wi )
= (dim m m
i=1 Vi )(dim i=1 Wi )
Ym
= dim Vi dim Wi
i=1
Ym
= dim Hom (Vi , Wi ).
i=1
So is a tensor map and
Hom (V1 , W1 ) Hom (Vm , Wm ) = Hom (m m

i=1 Vi , i=1 Wi ),
i.e., Hom (m m
i=1 Vi , i=1 Wi ) is a tensor product of Hom (V1 , W1 ), . . . , Hom (Vm , Wm ).
Similarly we can view CQm

i=1 ki
Qm
i=1 nm
as the tensor product of Ck1 n1 , . . . , Ckm nm
such that the Kronecker products A1 Am are the decomposable tensors.
Problems
1. Define the bilinear map : V V V V by (v1 , v2 ) = v2 v1 so
that there is T End (V V ) such that T = , i.e., T (v1 , v2 ) = v2 v1 .
Prove that when dim V = 1, T is not an induced map, i.e., there are no
T1 , T2 End (V ) such that T = T1 T2 .
2. Let {e1 , . . . , en }Pbe a P
basis for V . Define Tij End V such that Tij ek =
n n
kj ei . Let T = i=1 j=1 Tij Tji . Show that T (v1 v2 ) = v2 v1 for
all v1 , v2 V (cf. Theorem 3.7.1 and notice that the T in Problem 1 is a
linear combination of induced maps).
3. Prove that in the proof of Theorem 3.7.1
i
(a) (Daniel) {Tjp : j = 1, . . . , ki , q = 1, . . . , ni } is a basis of Hom (Vi , Wi ).

(b) {T : 2 , 1 } is a basis of Hom (m m
i=1 Vi , i=1 Wi ).
4. Let Vi be an inner product space, i = 1, . . . , m. Equip m i=1 Vi with the

inner product. Let T End (m V
i=1 i ). Prove that (T v
, v ) = 0 for all
v m V
i=1 i if and only if T = 0. (Hint: Show that (T u
, v ) = 0).
3.8. SOME MODELS OF TENSOR PRODUCTS 75
1.
2.
3.
4.
3.8 Some models of tensor products

Let M (V, . . . , V ; C) denote the space of all m-multilinear maps f : m V C.
Let E = {e1 , . . . , en } be a basis of V and let E = {f1 , . . . , fn } be the dual basis
of V , i.e.,
fi (ej ) = ij , i, j = 1, . . . , n.
Qm
Theorem 3.8.1. 1. { i=1 f(i) : m,n } is a basis of M (V, . . . , V ; C).
2. M (V, . . . , V ; C) = m V , i.e., there is tensor map : V V

M (V, . . . , V ; C) and dim M (V, . . . , V ; C) = nm .
3. M (V , . . . , V ; C) = m V , i.e., there is tensor map : V V

M (V , . . . , V ; C) and dim M (V , . . . , V ; C) = nm .
Qm
Proof. (1) We first show that S := { i=1 f(i) : m,n } spans M (V, . . . , V ; C).
First observe that for each e = (e(1) , . . . , e(m) ) V V and m,n ,
m
Y m
Y
( f(i) )e = f(i) (e(i) ) = . (3.8)
i=1 i=1
Let f M (V, . . . , V ; C). Then we claim
X m
Y
f= f (e ) f(i)
m,n i=1
where e = (e(1) , . . . , e(m) ). It is because from (3.8)

X m
Y X
f (e ) f(i) e = f (e ) = f (e ), m,n .
m,n i=1 m,n
We now show that S is a linearly independent set. Set
X m
Y
c f(i) = 0.
m,n i=1
Then
X m
Y
0= c f(i) (e ) = c , m,n .
m,n i=1
(2) It is easy to see that the map : m V M (V, . . . , V ; C) defined by

m
Y
(g1 , . . . , gm ) = gi
i=1
Qm
is multilinear. From (1) { i=1 f(i) : m,n } is a basis of M (V, . . . , V ; C).
So
dim M (V, . . . , V ; C) = |m,n | = nm = (dim V )m = (dim V )m .
So is a tensor map and hIm i = M (V, . . . , V ; C), i.e., M (V, . . . , V ; C) =
m V .
(3) Similar toQ(2). Indeed we can define : m V M (V , . . . , V ; C) by
m
(v1 , . . . , vm ) = i=1 vi (a notation) where
m
Y m
Y
( vi )(g1 , . . . , gm ) := gi (vi )
i=1 i=1
So M (V, . . . , V ; C) is a model for m V and M (V , . . . , V ; C) is a model

for m V .
We have another model for n V , namely M (V, . . . , V ; C) , i.e., the dual
space of M (V, . . . , V ; C).
Theorem 3.8.2. M (V, . . . , V ; C) is a model for m V , i.e., there is tensor map
: V V M (V, . . . , V ; C) and dim M (V, . . . , V ; C) = nm .
Qm
Proof. For v1 , . . . , vm V define i=1 vi M (V, . . . , V ; C) by
m
Y
( vi )f = f (v1 , . . . , vm ), f M (V, . . . , V ; C).
i=1
Qm
From Theorem 3.8.1(1),Qm { i=1 f(i) : m,n } is a basis of Q M (V, . . . , V ; C).
m
Now from (3.8) { i=1 e(i) : m,n } is the dual basis of { i=1 f(i) :

m,n } and thus is a basis of M (V, . . .Q , V ; C) . Then define : V , V
m
M (V, . . . , V ; C) by (v1 , . . . , vm ) = i=1 vi .
Elements in M (V, . . . , V ; C) are called contra-variant tensors; elements
in M (V , . . . , V ; C) are called covariant tensors. They are useful tools in
differential geometry.
The tensor space
p q
z }| { z }| {
Vqp = V V V V
3.8. SOME MODELS OF TENSOR PRODUCTS 77
is called a tensor space of type (p, q) (with covariant type of degree p and
with contra-variant type of degree q). Analogous to the previous treatment,
under some tensor map, M (V , . . . , V , V, . . . , V ; C) (p copies of V and q copies
of V ) is a model of Vqp :
Let E = {e1 , . . . , en } be a basis of V and let E = {f1 , . . . , fn } be the dual
basis of V . Then
p
Y q
Y
{ e(i) f(j) : p,n , q,n }
i=1 j=1
is a basis for
M (V , . . . , V , V, . . . , V ; C) (p copy of V, q copy of V )
Define : V V V V M (V , . . . , V , V, . . . , V ; C) by
p
Y q
Y
(e(1) , . . . , e(p) , f(1) , . . . , f(q) ) = e(i) f(j) .
i=1 j=1
{e(1) e(p) f(1) f(q) : p,n , q,n }

= {e
f : p,n , q,n }
is a basis of Vqp .
Problems
1. Define a simple tensor map : m V (m V ) such that m V =
(V ) .
2. Let M (V1 , . . . , Vm ; W ) be the set of all multilinear maps
Qm from V1 Vm
to W . Prove that dim M (V1 , . . . , Wm ; W ) = dim W i=1 dim Vi .

. . . , fm V . Define (f1 , . . . , fm ) = f (m V ) by f (v1
1. Let f1 ,Q
m
vm ) = i=1 fi (vi ). Clearly is multilinear.
2.
Chapter 4
Symmetry classes of tensors
4.1 Permutation operators

Let V be an n-dimensional inner product space. Let Sm be the symmetric group
of degree m on the set {1, . . . , m}. Each Sm yields a multilinear map
(v1 , . . . , vm ) = v1 (1) v1 (m) , v1 , . . . , vm V.
By the unique factorization property,

m VH / m V
HH v v
HH vv
H
HHH
v
vv
$ zvv P ()
m
V
there is unique P () End (m V ) such that P = , i.e.,
P ()(v1 v2 vm ) := v1 (1) v1 (2) v1 (m) ,
for all v1 , . . . , vm V . The operator P () called the permutation operator

associated with on m V .
E
Theorem 4.1.1. For each Sm , ([P ()]E )
,
= ,1 , where ,
m,n .
Proof. Let E = {e1 , . . . , em } be a basis of V and let E = {e : m,n }
which is a basis of m V . Then
P ()e
= P ()e(1) e(m)
= e(1 (1)) e(1 (m))
= e
1
X
= ,1 e
, m,n .
m,n
79
80 CHAPTER 4. SYMMETRY CLASSES OF TENSORS
So the matrix representation is

E
([P ()]E )
,
= (,1 )m,n
m,n
From the proof we see that
P ()e
= e 1 (4.1)
The permutation operator yields a representation of Sm , i.e., P : Sm GL(m

i=1 Vi ).
Theorem 4.1.2. Let , Sm . Then

(a) P () = P ()P ().
(b) P (e) = Im V .
(c) P () is invertible and P ()1 = P ( 1 ).
(d) P () is unitary, i.e., P ()1 = P () with respect to the induced inner
product of m V .
(e) If dim V 2, then P () = P () implies = , i.e., P is a faithful
representation.
(f) tr P () = nc() , where n = dim V and c() is the number of factors in the
disjoint cycle factorization of .
Proof. (a) From (4.1)
P ()P ()v = P ()v1 = v1 1 = v()

1 = P ()v

(b) Clear or from Theorem 4.1.1.

(c) From (a) and (b).
(d) With respect to the inner product
m
Y m
Y
(P ()u , v ) = (v1 (i) , vi ) = (ui , v(i) ) = (u , v ) = (u , P ( 1 )v ).
i=1 i=1
Since m V is spanned by the decomposable tensors, P ()1 = P ( 1 ) =

P () , i.e., P () is unitary for all Sm .
(e) Since dim V 2, there are linearly independent vectors e1 , e2 V . Set
vk := e1 + ke2 , k = 1, . . . , m.
When i 6= j, vi and vj are linearly independent. If P () = P (), then

v1 = v1 6= 0. By Theorem 3.3.3, v1 (i) and v1 (i) are linearly
dependent for all i. So 1 (i) = 1 (i) for all i, i.e., = .
4.1. PERMUTATION OPERATORS 81
(f) From Theorem 4.1.1

X X
tr P () = ,1 = , = |{ m,n : = }|.
m,n m,n
In other word, tr P () is the number of elements m,n that is fixed

by ( 7 ). Decompose into the product of c() disjoint cycles. It
is easy to see that for each cycle all the -components must be the same
and they can be chosen from 1, . . . , n. So tr P () = nc() .
We remark that when Sm is replaced by any subgroup G of Sm , it yields a

representation P : G GL(m V ).
Problems
1. Suppose m > 1. Prove that unless = e or dim V = 1, P () is not an
induced operator, i.e., there are no Ti End V such that P () = m
i=1 Ti .
2. Suppose that Sm is a transposition. Prove that if P ( ) is a sum of k

induced operators, then k n2 , where n = dim V .
3. If dim V > 1, then 7 P () is a reducible representation of Sm .

1. Suppose that there were Ti End V such that P () = m
i=1 Ti . For all
0 6= v V we have
T1 v Tm v = (m m m
i=1 Ti )(i=1 v) = P ()(i=1 v) = v v.
Qm
Thus by Theorem 3.3.2 and Theorem 3.3.3 we have Ti v = ci v, i=1 ci = 1
for all i = 1, . . . , m. Notice that ci apparently (but really not) depends
on v. choose an orthonormal basis E = {e1 , . . . en } of V (after equip V
with an inner product). Each Ti is unitary since P () is unitary (Problem
(j) (j)
3.6 #2). Now Ti ej = ci ej , for all i, j with |ci | = 1, and T (es + et ) =
c(es + et ) with |c| = 1 where c depends on Ti . Since n = dim V > 1,
(s) (t)
c(es + et ) = Ti (es + et ) = ci es + ci et ,
for any s 6= j. So cs = ct for all s 6= t, i.e., Ti = cI. Hence m
i=1 Ti is a
scalar multiple of the identity, contradiction.
2.
3. Each P () is a permutation matrix P in Cnm nm . If dim V > 1, then
nm > 1. Notice that the span of m,n e is an invariant subspace
under P () for all m,n . Hence by Theorem 2.2.2, P : G GL(m V )
is reducible.
4.2 Symmetric multilinear maps and symmetriz-

ers
A multilinear map : m V W is said to be completely symmetric if
(v(1) , . . . , v(m) ) = (v1 , . . . , vm ), for all Sm . (4.2)
Each multilinear map : m V W induces a completely symmetric multilin-

ear map:
1 X
1 (v1 , . . . , vm ) = (v(1) , . . . , v(m) ).
m!
Sm
It is because
1 (v(1) , . . . , v(m) )
1 X
= (v(1) , . . . , v(m) )
m!
Sm
1 X
= (v (1) , . . . , v (m) ) ( = )
m!
Sm
= 1 (v1 , . . . , vm ),
i.e., 1 is a completely symmetric multilinear map according to (4.2).

It is easy to see that (4.2) is equivalent to the following
1 X
(v(1) , . . . , v(m) ) = (v1 , . . . , vm ) for all v1 , . . . , vm V. (4.3)
m!
Sm
Similarly, is said to be skew symmetric if
(v(1) , . . . , v(m) ) = ()(v1 , . . . , vm ) for all and v1 , . . . , vm V. (4.4)
Notice that (4.4) is equivalent to the following
1 X
( 1 )(v(1) , . . . , v(m) ) = (v1 , . . . , vm ). (4.5)
m!
Sm
Each multilinear : m V W yields a skew symmetric multilinear :

m V W :
1 X
(v1 , . . . , vm ) = ( 1 )(v(1) , . . . , v(m) ).
m!
Sm
4.2. SYMMETRIC MULTILINEAR MAPS AND SYMMETRIZERS 83
It is because
(v(1) , . . . , v(m) )
1 X
= ( 1 )(v(1) , . . . , v(m) )
m!
Sm
1 X
= ( 1 )(v (1) , . . . , v (m) ) ( = )
m!
Sm
1 X
= ( 1 )()(v (1) , . . . , v (m) ) (by ( 1 ) = ( 1 )())
m!
Sm
= () (v1 , . . . , vm ).
When m = 2, can be decomposed as

1 1
(v1 , v2 ) = [(v1 , v2 ) + (v2 , v1 )] + [(v1 , v2 ) (v2 , v1 )]
2 2
= 1 (v1 , v2 ) + (v1 , v2 ), (4.6)
i.e., = 1 + .
When m > 2, the decomposition of may have more than two summands.
The reason for m = 2 case having two summands is that
I(S2 ) = {1, }
is the set of irreducible characters of S2 .

For general subgroup G of Sm , the set of irreducible characters I(G) of G
will come into the picture.
We say that the multilinear map : m V W is symmetric with
respect to G and if
(e) X
( 1 )(v(1) , . . . , v(m) ) = (v1 , . . . , vm ) (4.7)
|G|
G
for all v1 , . . . , vm V . If is linear, then : m V W is symmetric with

respect to G and if and only if (v(1) , . . . , v(m) ) = ()(v1 , . . . , vm ) for all
G (Problem 3). When is not linear, they are not equivalent.
Motivated by 1 and , each multilinear map : m V W induces a
mulitlinear map : m V W symmetric with respect to G and :
(e) X
(v1 , . . . , vm ) := ( 1 )(v(1) , . . . , v(m) )
|G|
G
for all v1 , . . . , vm V . The map is called the component of with

respect to G and .
Theorem 4.2.1. Let : m V W be multilinear. Then is symmetric
with respect to G and , i.e., = .
Proof. We now show that satisfies (4.7) by using the irreducible character
orthogonal relation of the first kind:
(v1 , . . . , vm )
(e) X
= ( 1 ) (v(1) , . . . , v(m) )
|G|
G
(e) X (e) X
= ( 1 ) ( 1 )(v(1) , . . . , v(m) )
|G| |G|
G G
!
2 X X
(e) 1 1
= ( )( ) (v (1) , . . . , v (m) ) ( = )
|G|2
G G

(e)2 X |G| 1
= ( ) (v (1) , . . . , v (m) ) (by Theorem 2.4.2)
|G|2 (e)
G
= (v1 , . . . , vm ). (4.8)
Clearly if : m V W is symmetric with respect to G and , then = ,

and vice versa.
Recall I(G) denotes the set of irreducible characters of G. Each : m V
W induces for all I(G). The following is an extension of (4.6).
P
Theorem 4.2.2. Let : m V W . Then = I(G) . In particular,
for : m V m V X
= (4.9)
I(G)
m m
where : V V .
Proof.
X
(v1 , . . . , vm )
I(G)
X (e) X
= ( 1 )(v(1) , . . . , v(m) )
|G|
I(G) G

X X
= 1 (e)( 1 ) (v(1) , . . . , v(m) )
|G|
G I(G)
X
= e,1 (v(1) , . . . , v(m) ) (by Corollary 2.4.15)
G
= (v1 , . . . , vm ). (4.10)
P
Since : m V m V is multilinear, = I(G) follows immediately.
Since : m V m V is multilinear, by the unique factorization property

of m V there is a unique T (G, ) End (m V ) such that
= T (G, ), (4.11)
i.e.,

m VH / m V
HH vv
HH vv
HH
HH vv
v
$ zvv T (G,)
m
V
and thus
(v1 , . . . , vm ) = T (G, )v ,
To determine T (G, ) consider
T (G, )v = (v1 , . . . , vm )
(e) X
= ( 1 ) (v(1) , . . . , v(m) )
|G|
G
(e) X
= ( 1 )v
|G|
G
(e) X
= ( 1 )P ( 1 )v
|G|
G
(e) X
= ()P ()v .
|G|
G
So
(e) X
T (G, ) = ()P () End (m V ) (4.12)
|G|
G
and is called the symmetrizer associated with G and since from (4.11),
it turns into which is symmetric to G and . We now see some basic
properties of T (G, ).
Theorem 4.2.3. Let , I(G) where G < Sm . Then
(a) T (G, ) is an orthogonal projection with respect to the induced inner

product on m V , i.e., T (G, )2 = T (G, ) and T (G, ) = T (G, ).
P
(b) I(G) T (G, ) = Im V .
(c) If 6= , then T (G, )T (G, ) = 0.

Proof. (a) From Theorem 4.1.2 P () = P ()P () so that

! !
2 (e) X (e) X
T (G, ) = ()P () ()P ()
|G| |G|
G G
2 (e) X X
= ()()P ()P ()
|G|2
G G
!
2 (e) X X
= ()( 1 ) P ( ) ( = )
|G|2
G G

2 (e) X |G|
= ( ) P ( ) (by Theorem 2.4.2)
|G|2 (e)
G
= T (G, ).
So T (G, ) is a projection. By Theorem 2.4.4 () = ( 1 ) so that
T (G, )
(e) X
= ()P ()
|G|
G
(e) X
= ( 1 )P ( 1 ) (P () = P ()1 by Theorem 4.1.2(d))
|G|
G
= T (G, ).
So T (G, ) is an orthogonal projection.
(b) By Theorem 2.4.15
X X (e) X
T (G, ) = ()P ()
|G|
I(G) I(G) G

X 1 X
= (e)() P ()
|G|
G I(G)
= P (e)
= Im V .
P
(c) By Theorem 2.4.3, G ()( 1 ) = 0 if 6= . So
T (G, )T (G, )
(e)(e) X X
= ()()P ()
|G|2
G G
!
(e)(e) X X 1
= ()( ) P ( )
|G|2
G G
= 0.
The range of T (G, )

Vm (G) := T (G, )(m V )
is called the symmetry class of tensors over V associated with G and .
Theorem 4.2.2 implies that
X
m V = V (G).
I(G)
The following asserts that the sum is indeed an orthogonal sum with respect to
the induced inner product on n V .
Theorem 4.2.4. The tensor space m V is an orthogonal sum of V (G),
I(G), i.e.,
m V =I(G) V (G).
In other words
m V = V1 (G) Vk (G),
where I(G) = {1 , . . . , k } is the set of irreducible characters of G.
Proof. From Theorem 4.2.3(a) and (c), T (G, ) is an orthogonal projection.
The elements in Vm (G) of the form
v1 vm := T (G, )(v1 vm )
are called decomposable symmetrized tensors. Such notation does not
reflect its dependence on G and . From Theorem 4.2.4, v1 vm is the
piece of v1 vm in V (G). So V (G) is spanned by the decomposable
symmetrized tensors v1 vm .
As we saw before, when G = S2 we have I(G) = {1, }. Then
1
v1 v2 = T (S2 , 1)(v1 v2 ) = (v1 v2 + v2 v1 )
2
and
1
v1 v2 = T (S2 , )(v1 v2 ) = (v1 v2 v2 v1 ).
2
Clearly
v1 v2 = v1 v2 + v1 v2 .
Moreover
(v1 v2 , v1 v2 )
1
= (v1 v2 + v2 v1 , v1 v2 v2 v1 )
4
1
= ((v1 v2 , v1 v2 ) (v1 v2 , v2 v1 ) + (v2 v1 , v1 v2 ) (v2 v1 , v2 v1 ))
4
1
= ((v1 , v1 )(v2 , v2 ) (v1 , v2 )(v2 , v1 ) + (v2 , v1 )(v1 , v2 ) (v2 , v2 )(v1 , v1 ))
4
= 0,
verifying that
V V = V1 (S2 ) V (S2 )
as indicated in Theorem 4.2.4.
The following is the unique factorization property of symmetrizer map :
m V V (G).
Theorem 4.2.5. Let : m V W be a multilinear map symmetric with
respect to G and . There exists a unique T Hom (V (G), W ) such that
= T ,
/ V (G)
m VE
EE x x
EE xx
E
EE
x
x T
" |xx
W
i.e,
(v1 , . . . , vm ) = T v .
Proof. Uniqueness follows immediately since the decomposable symmetrized
tensors v span V (G).
For the existence, from the factorization property of (Theorem 3.2.3),

m VF / m V
FF x x
FF x
F
FF xxxT
" x| x
W
there is T Hom (m V, W ) such that = T . Then from the symmetry of

with respect to G and ,
(v1 , . . . , vm ) = (v1 , . . . , vm )
(e) X
= ( 1 )(v(1) , . . . , v(m) )
|G|
G
(e) X
= ( 1 )T v ( = T )
|G|
G
!
(e) X
= T ( )P ( ) v
1 1
|G|
G
= T T (G, )v
= T v
Set T := T |V (G) Hom (V (G), W ). So
(v1 , . . . , vm ) = T v .
Problems
1. Prove that if G, then P ()T (G, ) = T (G, )P (). In addition, if
(e) = 1, then P ()T (G, ) = ( 1 )T (G, ). (Hint: If is linear, then
() = ()())
2. Prove P ( 1 )v = v . In addition, if (e) = 1, then v = ()v .
3. Show that if is linear, then : m V W is symmetric with respect
to G and if and only if (v(1) , . . . , v(m) ) = ()(v1 , . . . , vm ) for all
G.
P Qm
4. Let f1 , . . . , fm V . Prove that : m V C defined by = G () t=1 f(t)
is symmetric with respect to G and .
5. Prove that the multilinear map : m V C is skew symmetric with re-
spect to Sm if and only if vi = vj implies that (v1 , . . . , vm ) = 0 whenever
i 6= j. (Hint: is a product of transpositions).

1. Essentially from the proof of Theorem 4.2.3
!
(e) X
P ()T (G, ) = P () ()P ()
|G|
G
(e) X
= ()P () (4.13)
|G|
G
(e) X
= ( 1 )P ( ) ( = )
|G|
G
!
(e) X
= ( )P ( ) P () (( 1 ) = ( ))
|G|
G
= T (G, )P ().
In addition, if is linear, then from (4.13)
P ()T (G, )
(e) X
= ()P ()
|G|
G
(e) X
= ( 1 )P ( )
|G|
G
(e) X
= ( 1 ) ( )P ( ) (() = ()() since linear )
|G|
G
1
= ( )T (G, ).
2. By Problem 1,
P ( 1 )v = P ( 1 )T (G, )v = T (G, )P ( 1 )v = T (G, )v = v .
In addition, if is linear, then again by Problem 1,
v = P ( 1 )v = P ( 1 )T (G, )v = ()T (G, )v = ()v .
3. Suppose is linear so that it is a representation.
If (v(1) , . . . , v(m) ) = ()(v1 , . . . , vm ) for all G, then
(e) X
( 1 )(v(1) , . . . , v(m) )
|G|
G
(e) X
= ( 1 )()(v1 , . . . , vm )
|G|
G
= (e) (v1 , . . . , vm ) (( 1 )() = ( 1 ) = (e))

2
= (v1 , . . . , vm ) ( since (e) = 1)
(Answer to Alexs question: Indeed for general irreducible character , if

(v(1) , . . . , v(m) ) = ()
(e) (v1P
, . . . , vm ) for all G, then is symmetric
with respect to G and since G ( 1 )() = |G| by Theorem 2.4.2)
If is symmetric with respect to G and , then = T for some T

Hom (V (G), W ) by Theorem 4.2.5. Thus
(v(1) , . . . , v(m) ) = T (v(1) , . . . , v(m) )

= T v
= T (()v ) (by Problem 2)
= ()T (v )
= ()(v1 , . . . , vm )
4.3. BASIC PROPERTIES OF SYMMETRIZED TENSOR 91
4. (Roy)
(e) X
( 1 )(v(1) , . . . , v(m) )
|G|
G
m
!
(e) X 1
X Y
= ( ) () f(t) (v(t) )
|G| t=1
G G
m
(e) X Y
= ( 1 )( ) f (t) (v(t) ) ( = )
|G| t=1
, G
! m
(e) X X 1
Y
= ( )( ) f (t) (vt ) (( ) = ( ))
|G| t=1
G G
Y m
(e) X |G|
= ( ) f (t) (vt ) (Theorem 2.4.3)
|G| (e) t=1
G
X m
Y
= ( ) f (t) (vt )
G t=1
= (v1 , . . . , vm )
Therefore is symmetric with respect to G and .
5. From Problem 3, is skew symmetric if and only if (v(1) , . . . , v(m) ) =

()(v1 , . . . , vm ). Thus, if = (ij), then skew symmetric implies
that (v1 , . . . , vi , . . . , vj , . . . , vm ) = (v1 , . . . , vj , . . . , vi , . . . , vm ). So if
vi = vj , then (v1 , . . . , vm ) = 0. Conversely suppose that vi = vj implies
that (v1 , . . . , vm ) = 0 whenever i 6= j. Then
0 = (v1 , . . . , vi + vj , . . . , vi + vj , . . . , vm )
= (v1 , . . . , vi , . . . , vi , . . . , vm ) + (v1 , . . . , vj , . . . , vj , . . . , vm )
+(v1 , . . . , vi , . . . , vj , . . . , vm ) + (v1 , . . . , vj , . . . , vi , . . . , vm )
So (v1 , . . . , vi , . . . , vj , . . . , vm ) = (v1 , . . . , vj , . . . , vi , . . . , vm ). Each

Sm is a product of transpositions, i.e., = 1 2 k . So
(v(1) , . . . , v(m) ) = (v2 k (1) , . . . , v2 k (m) ) = = ()(v1 , . . . , vm ).
4.3 Basic properties of symmetrized tensor

The symmetrized tensors, i.e., the elements in V (G) is a subspace of m V
but the decomposable symmetrized tensors in V (G) may not be decomposable
elements in m V . When G = {e}, must be trivial and V (e) = m V so
m V can be viewed a special kind of V (G). So basic properties of V (G) apply
to m V as well.
Theorem 4.3.1. If u + + v = 0, then for any multilinear map : m V

W symmetric with respect to G and ,
(u1 , . . . , um ) + + (v1 , . . . , vm ) = 0.
Proof. From Theorem 4.2.5, there is T Hom (V (G), W ) such that
(u1 , . . . , um ) = T u , . . . , (v1 , . . . , vm ) = T v .
So
(u1 , . . . , um ) + + (v1 , . . . , vm ) = T (u + + v ) = 0.
Notice that when G = {e}, the symmetric multilinear function with respect
to {e} and 1 simply means that is simply multilinear. So Theorem 4.3.1
is a generalization of Theorem 3.3.1.
Theorem 4.3.2. If v1 vm = 0, then v1 , . . . , vm are linearly dependent.
Proof. Suppose on the contrary that v1 vm = 0 and v1 , . . . , vm were linearly
independent. Extend it to a basis {v1 , . . . , vm , vm+1 , . . . , vn } of V . So
{v(1) v(m) : G} {v(1) v(m) : m,n }
must be linearly independent. Now

(e) X
0 = v = ( 1 )v(1) v(m) .
|G|
G
So all the coefficients are zero, contradicting (e) = n 6= 0.

Necessary and sufficient condition for v1 vm = 0 is still unknown. See
[16, 17, 6] for some partial results.
The conclusion in Theorem 4.3.2 is weaker than Theorem 3.3.2.
The necessary and sufficient condition for u = v (6= 0) is not known for
general G and . When G = Sm (and its irreducible characters are called
immanent), necessary and sufficient conditions are given [25]. A necessary
condition for equality of immanantal decomposable tensors, unconstrained by
the families of vectors (u1 , . . . , um ) and (v1 , . . . , vm ), was given in [23].
The following is a necessary condition.
Theorem 4.3.3. If u = v 6= 0, then hu1 , . . . , um i = hv1 , . . . , vm i.
Proof. Set W := hv1 , . . . , vm i. Let {e1 , . . . , ek } (k m) be a basis of W .
Suppose on the contrary, ui 6 W for some i. Then extend {e1 , . . . , ek } to a
basis E = {e1 , . . . , ek , ek+1 = ui , ek+1 , . . . , en } of V and let {f1 , . . . , fn } be the
dual basis of V , i.e., fi (ej ) = ij . So
fr (vj ) = 0, if k < r n and j = 1, . . . , m (4.14)

fs (ui ) = 0, if 1 s k (4.15)
Qm
Now for each m,n , t=1 f(t) : m V C is multilinear. So by the unique
factorization property of , there is T (m V ) such that
m
Y
f(t) = T .
t=1
In particular
m
Y m
Y
T e
= T e = ( f(t) )e = f(t) (e(t) ) = .
t=1 t=1
In other words, {T : m,n } is a basis of (m V ) dual to E . From v 6= 0,

we claim that there is m,n such that T v 6= 0. Reason: if T v = 0 for
all , then write (since v m V )
X
v = a e

m,n
so that for all m,n

X X
0 = T v = a T e
= a = a ,
m,n m,n
contradicting that v 6= 0. Now

m
(e) X (e) X Y
0 6= T v = ( 1 )T v = ( 1 ) f(t) (v(t) ). (4.16)
|G| |G| t=1
G G
So m,k (otherwise some (t) > k so that f(t) (v(t) ) = 0 from (4.14)).
From (4.15),
m
(e) X Y
0 6= T v = T u = ( 1 ) f(t) (u(t) ) = 0,
|G| t=1
G
a contradiction. So u1 , . . . , um hv1 , . . . , vm i. By symmetry, hu1 , . . . , um i =

hv1 , . . . , vm i.
The conclusion in Theorem 4.3.3 is weaker than Theorem 3.3.3.
The inner product on V induces an inner product on m V . Such induced
inner product descends on V (G) via restriction. We now see the explicit form
of the induced inner product on V (G).
Theorem 4.3.4. Let V be an inner product space. Then the induced inner
product on V (G) is given by
m
(e) X Y
(u , v ) = (u , v ) = (u , v ) = () (ut , v(t) ).
|G| t=1
G
Proof. Since T (G, ) is an orthogonal projection Theorem 4.2.3(a) and ( 1 ) =

(),
(u , v ) = (T (G, )u , T (G, )v )
= (u , T (G, )v ) = (u , v ) = (u , v )
!
(e) X
= ()u 1 , v

|G|
G
(e) X
= ()(u
1 , v )
|G|
G
m
(e) X Y
= () (u1 (t) , vt ) (induced inner product)
|G| t=1
G
m
(e) X Y
= () (ut , v(t) ).
|G| t=1
G
Let
G := { G : = } < G
be the stabilizer of m,n .
Theorem 4.3.5. (a) Let E = {e1 , . . . , en } be an orthonormal basis of V .
Then
(e) X
(e , e ) = (), , , m,n .
|G|
G
(b)
(e) X
ke k2 = (), m,n .
|G|
G
Proof. (a) From Theorem 4.3.4

m
(e) X Y (e) X
(e , e ) = () (e(t) , e(t) ) = (), .
|G| i=1
|G|
G G
(b) It follows from (a) by setting = :

(e) X (e) X
ke k2 = (), = ().
|G| |G|
G G
Theorem 4.3.6. Let V be a vector space with dim V = n. Then

(e) X
dim V (G) = ()nc() .
|G|
G
Proof. Since T (G, ) is an orthogonal projection by Theorem 4.2.3(a),
dim V (G) = rank T (G, ) = tr T (G, )

(e) X
= ()tr P ()
|G|
G
(e) X
= ()nc() . (by Theorem 4.1.2(f))
|G|
G
Since dim m V = nm we have

X X (e) X
nm = dim m V = V (G) = ()nc()
|G|
I(G) I(G) G
which also follows from Theorem 2.4.15.
Problems
1. Let E = {e1 , . . . , en } be an orthonormal basis of V . Prove that if

2
Qm,n , then ke k2 = (e)|G| .
2. Let E = {e1 , . . . , en } be an orthonormal basis of V and m,n , G.

P
Prove that (e , e ) = (e)
|G| G (
1
).
3. Suppose m n. Prove that V (G) 6= 0.
4. Suppose
P is an irreducible character of G and m,n . Prove that
1
|G | G () is a nonnegative integer (Hint: consider the restriction
of onto G ).
Solutions to Problem 4.3
1. By Theorem 4.3.5(b) with G = e.
2. By Theorem 4.3.5(a),
(e) X (e) X (e) X

(e , e ) = (), = ( 1 ), = ( 1 ).
|G| |G| |G|
G G G
3.
4. which is a nonnegative integer.

4.4 Bases of symmetry classes of tensors

In the last section we have a formula for the dimension of V (G) and the ex-
pression
(e) X
dim V (G) = ()nc()
|G|
G
is not simple at all. In this section we want to construct a basis of V (G).

It involves two steps: (1) express V (G) as a direct sum of orbital subspaces
(Theorem 4.4.5 and (2) find bases for orbital subspaces. To this end we first
decompose m,n via G.
Let G < Sm . Recall
m,n = { = ((1), . . . , (m)), 1 (j) n, j = 1, . . . , m}.
Each G acts on m,n , i.e., : m,n m,n defined by
() = 1 , m,n .
Since
() = ( 1 ) = 1 1 = (), m,n
the action 7 is a a homomorphism (isomorphism if n > 1) from G into the
manifestation of Snm as a group of permutations of m,n , i.e.,
G = { : G}
is a group of permutations of m,n .
Two sequences and in m,n are said to be equivalent modulo G (more
precisely, modulo G), denoted by (mod G), if there exists G such
that = 1 () = , where := (((1)), . . . , ((m))). This equivalence
relation partitions m,n into equivalence classes.
For each m,n the equivalence class
= { : G}
is called the orbit containing . So we have the following disjoint union
m,n = .
Let be a system of representatives for the orbits such that each sequence in
is first in its orbit relative to the lexicographic order. Now
H < G G H
since G implies that for all G and thus for all H.
Consider the example 3,2 .
1. If G = S3 , then
3,2 = {(1, 1, 1), (1, 1, 2), (1, 2, 1), (1, 2, 2), (2, 1, 1), (2, 1, 2), (2, 2, 1), (2, 2, 2)}
has 4 orbits:
4.4. BASES OF SYMMETRY CLASSES OF TENSORS 97
(a) {(1, 1, 1)}.

(b) {(1, 1, 2), (1, 2, 1), (2, 1, 1)}.
(c) {(1, 2, 2), (2, 1, 2), (2, 2, 1)}.
(d) {(2, 2, 2)}.
Then = {(1, 1, 1), (1, 1, 2), (1, 2, 2), (2, 2, 2)}.
2. If G = {e, (12)}, then 3,2 has 6 orbits:
= {(1, 1, 1), (1, 1, 2), (1, 2, 1), (1, 2, 2), (2, 2, 1), (2, 2, 2)}.
We use the notations and < with respect to the lexicographical

order. The following theorem is clear.
Theorem 4.4.1. If m,n , then if and only if for all G.
Theorem 4.4.2. (a) Gm,n for all G < Sm .
(b) If G = Sm , then = Gm,n .
(c) If m n and = Gm,n , then G = Sm .
Proof. (a) If Gm,n , then clearly for all G. So .
(b) If , then there is Sm such that Gm,n . So =
Gm,n . Then apply (a).
(c) Suppose G 6= Sm . Then there is a transposition 6 G. Since m n, =
(1, . . . , m) Qm,n Gm,n . Now and are not in the same orbit (otherwise
there is G such that = ; but then = G, a contradiction).
In other words, 6= for any G. Moreover, since 6= e, we have
6 Gm,n for any G. Thus the representative (in ) of the orbit is
not contained in Gm,n , contradicting = Gm,n .
??? The conclusion in (c) is not valid if m > n, for example, 3,2 with
G = {e, (12)}.
We develop some combinatorial machinery in order to discuss the structure
of a basis for V (G).
Recall that the stabilizer G < G is a subgroup of G. For any G, recall
the right coset of G in G
G := { : G }
containing . Since G is a disjoint union of right cosets, i.e., there are 1 , . . . , k

G such that
G = G 1 G k .
Moreover |G i | = |G | so that k|G | = |G|. Clearly = {1 , . . . , k }
(1 , . . . , k depend on ).
The following is a combinatorial formula involving and m,n that we will
need in the construction of a basis for V (G).
Theorem 4.4.3. Let : m,n W be a map. Then

X X 1 X
() = ().
|G |
m,n G
Proof. Recall m,n = so that

X X X
() = ().
m,n
Write = {1 , . . . , k }. Notice that
|{ G : = j }| = |{ G : j1 = }| = |{ G : = }| = |G |
j = 1, . . . , k, i.e., when runs through G, meets j with |G | occurrences.

So
X Xk
1 X
() = (j ) = ().
j=1
|G |
G
Let E = {e1 , . . . , en } be a basis of V . Since E = {e

: m,n } is a basis
of m V and m,n = ,
X X
V (G) = he : m,n i = he : i = he : Gi. (4.17)

The subspace
O := he : Gi V (G)
is called the orbital subspace associated with m,n . The following gives
some relation between orbital spaces.
Theorem 4.4.4. Let E = {e1 , . . . , en } be an orthonormal basis of the inner

product space V . Then
(a) (e , e ) = 0 if 6 (mod G).
(b) ke k = ke k if (mod G).
Proof. (a) Notice that (e , e ) = 0 if 6 (mod G) by Theorem 4.3.5(a).

(b)
PBy Theorem 4.3.5(b)
P it suffices to show that if = for some G,
then G () = G ( ). Indeed G = 1 G by Problem 2. So
X X X X
( ) = (1 ) = () = ().
G G 1 G G
So for any m,n , e = 0 if and only if e = 0 for all G. In other

words, if one element in the orbital space O isP zero, then all elements in O
are zero. By Theorem 4.3.5 e 6= 0 if and only if G () 6= 0. So we define
X
:= { : () 6= 0} (4.18)
G
where G = { G : = } is the stabilizer (a subgroup of G) of . Notice

that
1 X
() = (|G , 1), (4.19)
|G |
G
the inner product (see Chapter 2) of the principal character and the character
|G (not necessarily irreducible), the restriction of onto G (it is a character
of G because the restriction of the representation A : G GLn (C) to G
is a representation). Notice that (|G , 1) is a nonnegative integer since it
denotes the multiplicity of the principal character in |G by Theorem 2.4.7
and Theorem 2.4.12.
Theorem 4.4.5. Let E = {e1 , . . . , en } be a basis of V . Then

V (G) =
he : Gi. (4.20)
Proof. Equip V with an inner product so that E is an orthonormal basis (The-

e = e = 0 for all G. So (4.17) becomes
orem 1.6.3). When \ ,
(4.20). By Theorem 4.4.4, the orbital subspaces are orthogonal, i.e., (4.20) is
an orthogonal sum and thus is a direct sum.
Notice that depends on m,n and G alone but depends on as well
(Compare
and is a very important set. We now see some basic properties of
Theorem 4.4.2).
Theorem 4.4.6. for all G < Sm (Qm,n = if m > n).
(a) Qm,n
= Qm,n .
(b) If G = Sm and = , then
= Qm,n , then G = Sm and = .
(c) If 1 < m n and
Proof. (a)
P By Theorem 4.4.2(a) Qm,n Gm,n . If Qm,n , then G = {e}
so that G () = (e) 6= 0. Thus by (4.20).
(b) Since G = Sm , by Theorem 4.4.2(b) = Gm,n . If \ Qm,n , then
there are i 6= j such that (i) = (j). Define the transposition = (ij). Then
G . Since ( ) = 1,
X X X
() = ( ) = ().
G G G
P So
Qm,n and hence = Qm,n by (a).
So G() = 0, i.e., 6 .
(c) FromP1 < m n, = (1, . . . , i 1, i, i, i + 2, . . . , m) Gm,n . Since 6

Qm,n = , G () = 0. Notice that (i, i + 1) G for all i = 1, . . . , m 1
(if
P the transposition (i, i + 1) 6 G, then we would have G = {e}, contradicting
G () = 0). Then G = Sm by Problem 2.1 #2.
Let := (i, i + 1) G. Then G = {e, } so that
X
(e) + ( ) = () = 0.
G
So ( ) = (e) = k where k is the degree of . Then from Problem 1,

A( ) = Ik if A is the (unitary) representation for . By Problem 2.1 #2 for
any G
= 1 r
where each j is a transposition of the form (i, i + 1). So
A() = A(1 r ) = A(1 ) A(r ) = (1)r Ik .
So
() = tr A() = (1)r k = (1 ) (r )k = k()
for all G, i.e., = k. Since is irreducible, we must have k = 1.
Let X
:= { m,n : () 6= 0}.
G
= and
Note that
=
=
{ : G}. (4.21)
So for m,n , e 6= 0 if and only if . So the set {e : } consists

of the nonzero elements of {e : m,n } (a spanning set of V (G)).
Moreover from (4.20)
V (G) = O , (4.22)
where O = he : Gi. In order to find a basis for V (G), it suffices to find
Define
bases of the orbital subspaces O , .
s := dim O = dimhe : Gi.
Notice that s depends on .
Theorem 4.4.7. (Freese)
(e) X
s = () = (e)(|G , 1).
|G |
G
Proof. Since G is a disjoint union of right cosets of G in G, i.e., there are

1 , . . . , k G such that
G = G 1 G k
and = {1 , . . . , k }. Moreover |G i | = |G | so that k|G | = |G|. Notice

that O = T (G, )W , where

W := he
: Gi = he : i.
Then E := {e
1 , . . . , ek } is a basis of W but {e1 , . . . , ek } may not be
a basis for O .
Since W is invariant under P () and thus invariant under T (G, ), the
restriction T (G, ) = T (G, )|W of T (G, ) is a linear map on W and is a
projection (Problem 1.8 #8).
Let C = (cij ) := [T (G, )]E E Ckk (C depends on and s := dim O

k). From
T (G, )W = T (G, )he

: Gi = he : Gi,
we have (by Theorem 1.8.2)
s = dimhe : Gi = rank T (G, ) = tr T (G, ) = tr C.
We now compute C:
T (G, )e
j = ej = T (G, )e
j
(e) X
= ()e
j 1
|G|
G
(e) X
= ( 1 j )e
( = j 1 )
|G|
G
k
(e) X X
= ( 1 j )e
(G = ki=1 G i )
|G| i=1
G i
k
X (e) X
= (i1 1 j )e
i ( = i )
i=1
|G|
G
for j = 1, . . . , k. So
(e) X
cij = (i1 j ), i, j = 1, . . . , k.
|G|
G
Since (i1 i ) = () for all i,
k
X k
X (e) X (e) X
s = tr C = cii = () = ().
i=1 i=1
|G| |G |
G G
Recall that (|G , 1) is a nonnegative integer when m,n . So s =

(e)(|G , 1) > 1 for each if is not linear. When is linear, s = 1
(Problem 6).
Let E = {e1 , . . . , en } be a basis of V . We now construct a basis of V (G).
For each , we find a basis of the orbital subspace O : choose a lexicograph-
ically ordered set {1 , . . . , s } from { : G} such that {e1 , . . . , es }
is a basis of O . Execute this procedure for each . If {, , } is the
lexicographically ordered set , take
= {1 , . . . , s ; 1 , . . . , s ; . . . }

to be ordered as indicated (often we pick 1 = ). Then {e : } is a basis

= {1 , 1 , . . . , } is in lexicographic order but note that the
of V (G) (Clearly
elements of need not be in lexicographical order, for example, it is possible
that 2 > 1 ). Such order (not unique) in is called an orbital order. Clearly
Although is not unique, it does not depend on the choice of basis E since ,
and C do not depend on E, i.e., if F = {f1 , . . . , fn } is another basis of V ,
is still a basis of V (G).
then {f : }
Example: Consider m = n = 3 and G = S3 . Recall that there are three
irreducible characters of S3 , namely, the principal character 1, the alternating
character and the character with degree 2:
(e) = 2, (23) = (12) = (13) = 0, (123) = (132) = 1.
We now construct , and for this degree two . Since G = Sm , = Gm,n

for 3,3 by Theorem 4.4.2 and the orbits (equivalence classes) are
{(1, 1, 1)};
{(1, 1, 2), (1, 2, 1), (2, 1, 1)};
{(1, 1, 3), (1, 3, 1), (3, 1, 1)};
{(1, 2, 2), (2, 1, 2), (2, 1, 1)};
{(1, 2, 3), (1, 3, 2), (2, 1, 3), (2, 3, 1), (3, 1, 2), (3, 2, 1)};
{(1, 3, 3), (3, 1, 3), (3, 3, 1)};
{(2, 2, 2)};
{(2, 2, 3), (2, 3, 2), (3, 2, 2)};
{(2, 3, 3), (3, 2, 3), (3, 3, 2)};
{(3, 3, 3)}.
Then by direct computation
= {(1, 1, 2), (1, 1, 3), (1, 2, 2), (1, 2, 3), (1, 3, 3), (2, 2, 3), (2, 3, 3)}.

For example, if = (1, 1, 1), then G = S3 and

X X
() = () = (e) + (123) + (132) = 2 1 1 = 0.
G G
If = (1, 1, 2), G = {e, (12)} so that

X
() = (e) + (12) = 2 + 0 = 2.
G
The orbit ( := (1, 2, 3)) | | = 6 and other orbits are of size 3.

First we consider the small orbits. For = (1, 1, 2), we have G =
{e, (12)} so that
(e) X
s = () = (e) + ((1 2)) = 2.
|G |
G
With 1 = e, 2 = (23), 3 = (13),
= {(1, 1, 2), (1, 2, 1), (2, 1, 1)} = {1 , 2 , 3 }.
From
(e) X 1
cij = (i1 j ) = ((i1 j ) + (i1 (12)j ))
|G | 3
G
we have
2
3 13 13
C= 1 2
13
3 3
13 13 2
3
Any two columns of C are linearly independent so that we can take 1 = =

(1, 1, 2), 2 = (1, 2, 1). The computation for the other five small orbits is
similar, i.e.,
(1, 1, 3), (1, 2, 2), (1, 3, 3, ), (2, 2, 3), (2, 3, 3)
Now we consider the larger orbit. For = (1, 2, 3), we have G = {e} so
that
(e) X
s = () = 2(e) = 4
|G |
G
With 1 = e, 2 = (23), 3 = (12), 4 = (123), 5 = (132), 6 = (13),
= {(1, 2, 3), (1, 3, 2), (2, 1, 3), (2, 3, 1), (3, 1, 2), (3, 2, 1)} = {1 , . . . , 6 }.
From
(e) X 1
cij = (i1 j ) = (i1 j ))
|G | 3
G
we have
2
3 0 0 13 31 0
0 2
31 0 0 13
3
0 13 2
0 0 13
C=
1
3
13 0 0 2
3 13 0
0 0 13 2
0
3 3
0 13 31 0 0 2
3
Notice that rank C = 4 (for example the columns 1, 2, 3, 4 or columns 1, 2, 3, 5
are linearly independent. So we can take 1 = = (1, 2, 3), 2 = (1, 3, 2),
3 = (2, 1, 3) and 4 = (2, 3, 1). Hence
=
{(1, 1, 2), (1, 2, 1);
(1, 1, 3), (1, 3, 1);
(1, 2, 2), (2, 1, 2);
(1, 2, 3), (1, 3, 2), (2, 1, 3), (2, 3, 1);
(1, 3, 3), (3, 1, 3);
(2, 2, 3), (2, 3, 2);
(2, 3, 3), (3, 2, 3)}
in which the order is not lexicographic.
It is known (Problem 6) that = if and only if is linear. In such

cases, {e : } is an orthogonal basis of V (G) if E = {e1 , . . . , en } is an
such that {e : }
orthogonal basis of V . If (e) > 1, can we choose
is an orthogonal basis for V (G)? When G = Dm (the diheral group) such
exists for every I(G) if and only if m is a power of 2 (Wang and Gong
(1991) [34, 35], Holmes and Tam (1992) [4]). On the other hand, every doubly
transitive subgroup G < Sm has an irreducible character for which no such
exists (Holmes (1995) [3])
We give several common examples of symmetry classes of tensors and in-
duced operators.
Example 4.4.8. Assume 1 m n, G = Sm , and = . Then V (G) is
the mth exterior space m V , = = Qm,n , the set of strictly increasing
sequences in m,n , = Gm,n , the set of nondecreasing sequences in m,n . See
Problem 6 and Theorem 4.4.6(b).
Example 4.4.9. Assume G = Sm and 1, the principal character. Then
=
V1 (G) is the mth completely symmetric space m V , = = Gm,n . See
Problem 7.
Example 4.4.10. Assume G = {e} < Sm ( 1). Then V (G) = m V ,
=
= = m,n . See Problem 7.
We know that from Theorem 4.3.6
= (e) X
dim V (G) = || ()nc() .
|G|
G
in terms of s .
The following gives another expression for ||
Theorem 4.4.11.
X X
dim V (G) = =
|| s = s

X (e) X
= ()
|G |
G
(e) X X
= ().
|G|
m,n G
Proof. It suffices to establish the last equality. By the proofs of Theorem 4.4.3
and Theorem 4.4.4
X X X 1 X X
() = ()
|G |
m,n G G G
X 1 X X
= () (G = 1 G )
|G |
G G
X 1 X
= |G| ().

|G |
G
Problems
1. If U Cnn is unitary and tr U = n, prove that U = In .
2. Let G < Sm , m,n . Prove that G = 1 G .

P |G| P
3. Prove that |G |
= |m,n | and || = m,n |G |
|G| .
4. Let {e1 , . . . , en } be a basis of V . Prove that {e : } is linearly

independent.
P (e) P
5. Write s () = s = |G | G () to indicate its dependence
on , an irreducible character of G < Sm and m,n .
(a) s () divides (e) for all I(G).

P |G|
(b) I(G) s () = |G | .
|G|
(c) Unless G () = G and 1, we have s < |G | .
(d) For each Sn and G, s () = s (). Here = ( ((1)), . . . , ((m)))

m,n .
(e) if , then (e) s () (e)2 . (Hint: Problem 2.4 #1).

=
6. Prove that if and only if (e) = 1.
(or = )
7. Prove that = if and only if 1.
8. Let . Prove that if (e) = 1, then for any G , () = 1, i.e.,

|G 1.
n 1
P
c() n+m1
1
P c()
9. Prove the identities: m = m! Sm ()n , m = m! Sm n .

10. Suppose V (G) 6= 0. Prove that if G 6= Sm or 6= , then there is
such that (i) = (j) for some i 6= j.
11. Let C Ckk be the matrix in the proof of Theorem 4.4.7, i.e., cij =
(e) P 1 2
|G| G (i j ). Give a direct proof of C = C.
Solution to Problems 4.4

1. Since the diagonal entries uii satisfies |uii |P
1 since UP Cnn is unitary.
n n
So n = tr U implies that n = |tr U | = i=1 uii i=1 |uii | = n. By
the equality case of the triangle inequality, uii = n and hence U = In
since U has orthonormal rows (columns).
2. G = 1 = 1 G , i.e., 1 G .
3.
is
he : Gi. So {e : }
4. By Theorem 4.4.5 V (G) =
linearly independent.
5.
6. From Theorem 4.4.7 and the definition of , = 1 = s (=

(e)(|G , 1)) for all . Then notice that (|G , 1) is a positive

integer for each .
7.
P
8. Suppose 1. For any , G () = |G | = 6 so that .
= . Then use Problem 6 to have
Hence = since (e) = 1.
= , i.e., 1
P
Conversely suppose () is a positive integer for
|G | G
all . So X X
| ()| |()| |G |.
G G
But for all G, |()| (e) = 1 by Problem 2.4 #1. So by the

triangle inequality, () = 1 for all G , i.e., |G 1. Pick =
to have the desired result since G = G.
(1, . . . , 1) =
9. Since the alternating character and the principal character 1 of Sm are

linear, By Problem 6, = for = , 1. So on one hand

n
dim V (Sm ) = || = |Qm,n | = ,
m

n+m1
dim V1 (Sm ) = || = |Gm,n | = (by Problem 7).
m
On the other hand, by Theorem 4.3.6
1 X
dim V (Sm ) = ()nc()
m!
Sm
and
1 X c()
dim V1 (Sm ) = n .
m!
Sm
is nonempty. Assume that all

10. Suppose V (G) 6= 0 so that have
distinct entries. Then m n evidently. Suppose G = Sm and we need
to show that 6= . Since G = Sm , = Gm,n (Theorem 4.4.2(b)) we
have Qm,n . By Theorem 4.4.6(a) = Qm,n and then by Theorem
4.4.6(c) = .
11. Remark: In the proof of Theorem 4.4.7 we know that C = (cij ) :=
[T (G, )]E
E Ckk is a projection matrix since T (G, ) End (W )

is a projection.
k
X
(C 2 )ij = cis csj
s=1
k
(e)2 X X X
= 2
(i1 s ) (s1 j )
|G| s=1
G G
2 k
X X X
(e)
= (i1 s ) (s1 1 j ) ( = 1 )
|G|2 s=1 G G
k
2 X X
(e)
= (i1 s )(s1 j )
|G|2 s=1 ,G
(e) X X
2
= (i1 )( 1 j ) ( = s )
|G|2
G G
!
(e)2 X X 1 1
= ()( i j )
|G|2
G G
(e) 2 X ( 1 j )
i
= by Theorem 2.4.3
|G| (e)
G
= cij .
So C is a projection matrix.
4.5 Linear operators on V (G)

Let T End V . Similar to the induced operator m T End (m V ), we want to
study the induced operator on V (G) induced by T . The map : m V V (G)
defined by
(v1 , . . . , vm ) = T v1 T vm
is multilinear and symmetric with respect to G and . By Theorem 4.2.5, there
is a unique K(T ) End (V (G))
/ V (G)
m VH
HH vv
HH vv
HH vv
HH v
# zvv K(T )
V (G)
satisfying
K(T )v1 vm = T v1 T vm .
Such K(T ) is called the induced operator of T on V (G). Notice that K(T )
depends on G, , m, n. We now discuss some basic properties of K(T ).
Theorem 4.5.1. Let T End (V ). Then V (G) is an invariant subspace of

m T and
K(T ) = m T |V (G) .
So K(T )v = (m T )v for all v V (G).
Proof. Notice that

(m T )P () = P ()(m T )
so that (m T )T (G, ) = T (G, )(m T ) (Problem 1). So V (G) is an invariant
subspace of (m T ). Thus (m T ) induces a linear operator on V (G) which is
K(T ), because
K(T )v = T v1 T vm = T (G, )T v1 T vm
= T (G, )(m T )v = (m T )T (G, )v
= (m T )v
and the decomposable v span V (G), i.e., K(T ) = m T |V (G) .
Theorem 4.5.2. Let S, T End (V ). Then
(a) K(IV ) = IV (G) .
(b) K(ST ) = K(S)K(T ). So T 7 K(T ) is a representation of GL(V ) in

GL(V (G)).
4.5. LINEAR OPERATORS ON V (G) 109
m,r |, where r = rank T .

(c) rank K(T ) = |
Proof. (a) Clear.
(b)
K(ST ) = m (ST )|V (G)

= (m S m T )|V (G) by Theorem 3.4.1
= m S|V (G) m T |V (G) by Theorem 4.5.1
= K(S)K(T ).
(c) Since r = rank T , there is a basis {v1 , . . . , vn } of V such that T v1 , . . . , T vr

are linearly independent and T vr+1 = = T vn = 0. Let ei = T vi , i = 1, . . . , r,
and extend them to a basis E = {e1 , . . . , er , er+1 , . . . , en } of V . Now

E = {e : }
and {v : }
are bases of V (G). If m,r , then
K(T )v = T v(1) T v(m) = e .
So
m,r }
{e = K(T )v :
is a subset of the basis E of V (G) so that it is a linearly independent set.
When 6 m,r , there is some i such that (i) > r. Then T v(i) = 0 so

that K(T )v = 0. So rank K(T ) = | m,r |.
Theorem 4.5.3. Let V be an inner product space and T End V . Equip

V (G) with the induced inner product.
1. K(T ) = K(T ).
2. If T is normal, Hermitian, positive definite, positive semidefinite, unitary,
so is K(T ).
Proof. (a) From Theorem 4.5.1 V (G) is invariant under m T and m T . So
K(T ) = (m T )|V (G)

= (m T ) |V (G) by Theorem 3.4.3
= (m T |V (G) ) Theorem 1.7.1(a)
= K(T ) .
(b) Since m T has the corresponding property by Theorem 3.6.2, by Theo-

rem 1.7.1(b) K(T ) = n T |V (G) has the corresponding property.
Theorem 4.5.4. Let E = {e1 , . . . , en } be a basis of V and T End V . Suppose
that A = [T ]E E is upper triangular with diag A = Q (1 , . . . , n ). Then B =
[K(T )]E is upper triangular with diagonal entries
m
t=1 (t) , , where

E
(in the orbital order).
E = {e : }
Proof. Since A = [T ]E
E is upper triangular,
X
T ej = j ej + aij ei , j = 1, . . . , n,
i<j
there is G such
where 1 , . . . , n are the eigenvalues of T . For each ,

that = 1 . Then
K(T )e = T e(1) T e(m)

X X
= ((1) e(1) + ai(1) ei ) ((m) e(m) + ai(m) ei )
i<(1) i<(m)
m
Y X
= ( (t) )e + c e .
t=1 S
where
S := { m,n : (i) (i), i = 1, . . . , m, with at least one strict inequality}.
We remark that some S may not be in . Fix S and we may

Clearly
assume that e 6= 0 (note: from (4.21) but may not be in ).

1 = for some G. From the definition of S ,
1 < = 1 .
In other words 1 comes strictly before 1 in lexicographic order.

then < in the orbital order, according to the construction
Case 1. If ,

of .
then e O = he : G} is a linear combination of those
Case 2. If 6 ,

e in the orbital subspace O , where . Each < in the
orbital order by Case 1.
So [K(T )]E
E is upper triangular, where E = {e : } in the orbital order.

Qm
In Theorem 4.5.5, the eigenvalues of K(T ) are evidently i=1 (i) ,
counting multiplicity. However the order of these eigenvalues apparently
,
depends on the order of 1 , . . . , n ; but it is merely a deception.
Theorem 4.5.5. Let 1 , . . . , n and s1 , . . . , sn be the eigenvalues and the sin-
gular values of T End V , respectively. Then
Qm counting multiplicities.
1. The eigenvalues of K(T ) are i=1 (i) , ,
Moreover for any Sn ,
m
Y
(i) ,

i=1
are the eigenvalues of K(T ), counting multiplicity. Hence f (1 , . . . , n ) :=

tr K(T ) is a symmetric function of 1 , . . . , n , i.e.,
f (1 , . . . , n ) = f ( (1) , . . . , (n) ), Sn .
Qm counting multiplicities.
2. The singular values of K(T ) are i=1 s(i) , ,
Moreover for any Sn ,
m
Y
s (i) ,

i=1
are the singular values of K(T ), counting multiplicities.

Proof. (a) The first part follows from Theorem 4.5.4. Schurs triangular theorem
or Jordan form allows any prescribed ordering of 1 , . . . , n . In other words, for
any Sn ,
Ym

(i) ,
i=1
are the eigenvalues of K(T ), counting multiplicities.

(b) Apply (a) on A A.
Qm Qn m ()
In order to compute det K(T ) we rewrite i=1 (i) as t=1 t t , where
mt () denotes the multiplicity of the integer t, i.e., the number of times t
appearing in . Clearly for all m,n ,
m
Y n
Y mt ()
(i) = t .
i=1 t=1
Theorem 4.5.6. Let T End V . Then

m
det K(T ) = (det T ) n || (4.23)
Proof. Since the determinant is the product of eigenvalues, by Theorem 4.5.5

n
Y Y n Y
Y n
Y
mt () mt ()
det K(T ) = t = t = qt t ,
t=1
t=1
t=1
P
where qt := mt () and mt () is the multiplicity of t in . In other words,
qt is the total number of times the integer t appearing among all the sequences
By Theorem 4.5.5(a)
in .
n
Y n
Y n
Y q
qt t = qt(t) =
1 (t)
t
t=1 t=1 t=1
which is true for all scalars 1 , . . . , n (view

Pn them as indeterminants). So qt = q,
a constant, for all t = 1, . . . , n. Since t=1 mt () = m,
n
X n X
X n
XX
nq = qt = mt () =
mt () = m||.
t=1 t=1
t=1

So
m
q= ||.
n
Then
n
n
!m
n ||
Y m Y
n ||
m
det K(T ) = t = t = (det T ) n || .
t=1 t=1
The following is an alternate approach for qt = q for all t:

From Problem 4.4 #2 (G = 1 G ) and #5 for any m,n , G
we have
|G | = |G |, s = s , mt () = mt ().
For any arbitrary Sn ,
G = G , s = s , mt ( ) = m 1 (t) ().
Moreover, as runs over m,n , runs over m,n as well. Then
X X X
qt = mt () = s mt () =
s mt () (s = 0 if 6 )

1 X 1 X
= |G |s mt ()
|G| |G |
G
1 X 1 X
= |G |s mt () (|G | = |G |, s = s , mt () = mt ())
|G| |G |
G
1 X
= |G |s mt () (Theorem 4.4.3, () = |G |s mt ())
|G|
m,n
Then for any Sn

1 X
qt = |G |s mt ( ) ( m,n = m,n )
|G|
m,n
1 X
= |G |s m 1 (t) () (G = G , s = s , mt ( ) = m 1 (t) ())
|G|
m,n
= q 1 (t) .
Theorem 4.5.7. Let E = {e1 , . . . , en } be an orthonormal basis of the inner
product space V and equip V (G) with the induced inner product. Let T
End V and A = (aij ) = [T ]E
E . Then for , m,n ,
m
(e) X Y
(K(T )e , e ) = () a(t)(t) . (4.24)
|G| t=1
G
Proof. Notice that K(T )e = T e(1) T e(1) . Since E is an orthonormal

basis,
(T e(t) , e(t) ) = a((t))(t) .
Then use Theorem 4.3.4 with ut = T e(t) and vt = e(t) :
m m
(e) X Y (e) X Y
(K(T )e , e ) = () (T e(t) , e(t) ) = () a(t)(t) .
|G| t=1
|G| t=1
G G
We now discuss the matrix representation of K(T ). Let E = {e1 , . . . , en } be

an orthonormal basis of the inner product space V and equip V (G) with the
induced inner product. From the construction of the basis E = {e : },
different orbit subspaces are orthogonal (Theorem 4.3.5). However the vectors
within the same orbit subspace may not be orthogonal. So the right side of
(4.24) is not necessarily the corresponding element of K(T ) with respect to the
basis E of V (G).
When (e) = 1, ( = by Problem 4.4 #6) E = {e : } is an
orthogonal basis of V (G) (but not necessarily orthonormal). Upon normaliza-
tion,
(s )
0 |G|
E = e :
|G |
is an orthonormal basis of V (G) (Problem 6). Thus if (e) = 1, then the matrix
representation of K(T ) with respect to E0 follows from the following theorem.
Theorem
q 4.5.8. Suppose (e) = 1 and E = {e1 , . . . , en } is a basis of V (so
E = { G|G|
0 is a basis of V (G) according to the lexicographic
e : }
|
ordering). If T End V and A = [T ]E

E , then
m
!
E0 1 X Y
[K(T )]E0 = p () a(t)(t) (4.25)
|G ||G | G t=1
,
Proof. Equip V with an inner product so that E is an orthonormal basis (The-

orem 1.6.3). Then E0 is an orthonormal basis of V (G) with respect to the
induced inner product. Since (e) = 1, use (4.24) to have
s s
X Ym
|G| |G| 1
(K(T ) e , e ) = p () a(t)(t)
|G | |G | |G ||G | G t=1
When is linear, similar to the Kronecker product of matrices, we define

the induced matrix K(A) using (4.25). Let (e) = 1 and let A Cnn . The
matrix K(A) C|||

defined by
|
X Ym
1
K(A), = p () a(t),(t) , , (4.26)
|G ||G | G t=1
is called the induced matrix. If T and A are as in Theorem 4.5.8, then

E0
[K(T )]E0 = K(A).
Since the induced matrix K(A) is the matrix representation of K(T ), we have
the following properties.
Theorem 4.5.9. Let A Cnn with eigenvalues 1 , . . . , n and singular values
s1 sn 0. Suppose is linear and let K(A) be the induced matrix of
A. Then
(a) K(I) = I|| = 1 P
, where ||
c()
|G| G ()n .
(b) K(AB) = K(A)K(B), where A, B Cnn .

(c) K(A1 ) = K(A)1 if A Cnn is invertible.
m,r |, where r = rank A.
(d) rank K(A) = |
(e) If A is upper triangular, then K(A) is also upper triangular.
Qm
(f) The eigenvalues
Qm of K(A) are i=1 (i) , . The singular values of
K(A) are i=1 s(i) , .
P Qm P 1
P Qm
(g) tr K(A) = t=1 (t) = |G | G () t=1 a(t),(t) .
m
(h) det K(A) = (det A) n || .
(i) K(A ) = K(A) .
(j) If A is normal, Hermitian, psd, pd, or unitary, so is K(A).
For (e) > 1, see [18] for the construction of orthonormal basis of V (G),
the matrix representation of K(T ), where T End V , and the induced matrix
K(A).
We now provide an example of K(T ) with nonlinear irreducible character.
Assume that G = S3 and is the degree 2 irreducible character. We have
(e) = 2, (12) = (23) = (13) = 0, (123) = (132) = 1.
Assume dim V = n = 2. Then by direct computation,
3,2 = {(1, 1, 1), (1, 1, 2), (1, 2, 1), (1, 2, 2), (2, 1, 1), (2, 1, 2), (2, 2, 1), (2, 2, 2)}
= {(1, 1, 1), (1, 1, 2), (1, 2, 2), (2, 2, 2)}
=
{(1, 1, 2), (1, 2, 2)}
=
{(1, 1, 2), (1, 2, 1), (1, 2, 2), (2, 1, 2)}.
Let E = {e1 , e2 } be a basis of V . Then
= {e
E = {e : }
(1,1,2) , e(1,2,1) , e(1,2,2) , e(2,1,2) }
is a basis of V (G). Observe that E is not an orthogonal basis even if E is an

orthonormal basis, since
1
(e(1,1,2) , e(1,2,1) ) = (e(1,2,2) , e(2,1,2) ) = .
3
For example, by Theorem 4.3.5
(e) X 2 1
(e(1,1,2) , e(1,2,1) ) = ()(1,1,2),(1,2,1) = ((23) + (123)) = .
|S3 | 6 3
S3
Moreover (check by working out the C matrices!)
e(2,1,1) = e(1,1,2) e(1,2,1) , e(2,2,1) = e(1,2,2) e(2,1,2) . (4.27)
Let T End V be defined by

a b
[T ]E
E = .
c d
By direct computation,

a2 d abc 0 abd b2 c 0
0 a d abc abd b2 c
2
b2 c abd
[K(T )]E
=
acd bc 2

E 0 ad2 bcd 0
2 2
acd bc bc acd 0 ad2 bcd
For example
K(T )e(1,1,2) = T e1 T e1 T e2
= (ae1 + ce2 ) (ae1 + ce2 ) (be1 + de2 )
= a2 be(1,1,1) + a2 de(1,1,2) + acbe(1,2,1) + acde(1,2,2)
+cabe(2,1,1) + cade(2,1,2) + c2 be(2,2,1) + c2 be(2,2,2)
= a2 de(1,1,2) + acbe(1,2,1) + acde(1,2,2)
+cabe(2,1,1) + cade(2,1,2) + c2 be(2,2,1) by (4.26)
2
= (a d abc)e(1,1,2) + 0e(1,2,1) + (acd bc2
)e(1,2,2) + (acd bc2 )e(2,1,2) .
The computations for K(T )e(1,2,1) , K(T )e(1,2,2) , K(T )e(2,1,2) are similar.
Furthermore, one can define the derivation operator DK (T ) of T by
d
DK (T ) = K(I + tT )t=0 ,
dt
which acts on V (G) in the following way:

m
X
DK (T )v1 vm = v1 vj1 T vj vj+1 vm .
j=1
Clearly T 7 DK (T ) is linear. With respect to the above example

2a + d 0 b 0
0 2a + d b b
[DK (T )]E .
E =

c 0 a + 2d 0
c c 0 a + 2d
Research problem: Given the Jordan structure of T Cnn (T End V ),

what is the Jordan structure of Cm (T ) (G = Sm and = ) where 1 m n?
The answer is given by Atkin [1]; also see Littlewood [13] for some simplified
treatment. Atkin also studied the problem for Kronecker product. In general
the corresponding problem for K(T ) is not known.
Problems
1. Prove that (m T )P () = P ()(m T ) and (m T )T (G, ) = T (G, )(m T ).
2. If V (G) 6= 0, prove that K(T ) is invertible if and only if T is invertible.
Moreover, K(T 1 ) = K(T )1 .
3. Let S, T End V are psd. Show that K(S + T ) K(S) + K(T ), i.e.,
K(S + T ) K(S) + K(T ) is psd. Hence S T implies K(S) K(T ).
(Hint: Use Problem 3.6 #2)
4. Prove that if G = Sm , = , rank T < m, then K(T ) = 0.
5. Suppose T End V is positive semidefinte with eigenvalues 1
n . Prove that K(T ) on V1 (G) has largest eigenvalue m
1 and smallest
eigenvalues m
n .
6. Suppose that (e) = 1 and E = {e1 , .q
. . , en } is an orthonormal basis of the
inner product V . Prove that E = { |G| e : }
0 is an orthonormal
|G |
basis for V (G).
7. Suppose that (e) = 1 and E = {e1 , . . . , en } is a basis of V . Let T End V
and [T ]E
E = A = (aij ). Prove that
m
1 X Y

([K(T )]E
E ), =

() a(t)(t) , , .
|G | i=1
G
P
8. Let P
m,n and be linear. Prove that if G () = 0, i.e., 6 ,
then G ()() = 0, where : m,n W is any map.
9. Suppose that is linear. If A, B Cnn , deduce directly from (4.26) that

K(AB) = K(A)K(B) (Hint: Compare Theorem 1.10.1).
P Qm
10. Show that tr K(A) = (e) (|G , 1) t=1 (t) .

1. (m T )P ()v = (m T )v1 = T v1 (1) T v1 (m) = P ()(m T )v .
P
Then (m T )T (G, ) = T (G, )(m T ) follows since T (G, ) = (e)|G| G ()P ()
is a linear combination of P (), G.
6= 0. Then use Theorem 4.5.6 since a matrix
2. Suppose V (G) 6= 0, i.e., ||
is invertible if and only if its determinant is nonzero.
3. By Problem 3.6 #4, m (S + T ) m S + m T So
K(S + T ) = m (S + T )|V (G) (m S + m T )|V (G)
= m S|V (G) + m T |V (G) = K(S) + K(T ).
So if S T , then S = T + R where R 0. Then K(R) 0 by Theorem
4.5.3 and hence
K(S) = K(T + R) K(T ) + K(R) K(T ).
4. Let r = rank T .
rank K(T ) (Theorem 4.5.2)
= |m,r |
= |m,r Qm,n | (Theorem 4.4.6 since G = Sm , = )
= 0
since all elements of m,r has some repeated entries as r < m so that
m,r Qm,n = .
5. By Theorem 4.5.5, since = m,n , the eigenvalues are Qm (i) ,
i=1
m,n . Since 1 n 0, the largest eigenvalue of K(T ) is m
1 and
the smallest eigenvalue is mn .
6. By Theorem 4.3.5(a), E0 is an orthogonal basis. It remains to show that
each vector in E0 has unit length. By Theorem 4.3.5(b)
s
|G| 2 |G| 2 (e) X 1 X

e = ke k = () = () = (|G , 1)
|G | |G | |G | |G |
G G
which is a positive integer since . Notice that

1 X 1 X
() |()| 1
|G | |G |
G G
since each |()| 1 ( is linear). So (|G , 1) = 1.

7.
8.
9.
Qm Qm
10. Notice that t=1 (t) = t=1 (t) for all G, s = (e)(|G , 1)
and G = 1 G , for all m,n and G. So
m
XY X m
Y
tr K(T ) = (t) = (e) (|G , 1) (t)
t=1

t=1
X m
Y
= (e) (|G , 1) (t) .
t=1
4.6 Generalized matrix functions

Let G < Sm , I(G) and let A Cmm . The function
X m
Y
dG (A) = () at(t) (4.28)
G t=1
is called the generalized matrix function association with G and .

Example:
(a) When G = Sm and = ,
X m
Y
dG (A) = det A := () at(t) .
Sm t=1
(b) When G = Sm and 1,

m
X Y
dG (A) = per A := at(t)
Sm t=1
is called the permanent of A.

(c) When G = {e} < Sm , then
m
Y
dG (A) = h(A) := att
t=1
which is the product of the diagonal entries.

Theorem 4.6.1. Let A Cmm . Then
(a) dG (Im ) = (e).
4.6. GENERALIZED MATRIX FUNCTIONS 119
(b) If A is upper triangular, then
dG (A) = (e) det A = (e)per A = (e)h(A).
(c) dG (AT ) = dG (A) (() := () for all G).
(d) dG (A ) = dG (A). So if A is Hermitian, then dG (A) is real.
Proof. (a) and (b) are trivial.

(c)
X m
Y X m
Y
dG (AT ) = () a(t)t = () at(t) = dG (A).
G t=1 G t=1
(d) dG (A ) = dG ((A)T ) = dG (A) = dG (A).
Theorem 4.3.4 and Theorem 4.5.7 can be restated respectively using gener-
alized matrix function as in the next two theorems.
Theorem 4.6.2. Let v1 , . . . , vm , u1 , . . . , um V and let A Cmm where

aij := (ui , vj ). Then for each , m,n ,
m
(e) X Y
(u , v ) = () (ut , v(t) ) (Theorem 4.3.4)
|G| t=1
G
(e) (e)
= d (A) = d ((ui , vj ))
|G| G |G| G
Theorem 4.6.3. Let E = {e1 , . . . , en } be an orthonormal basis of V and let

T End V . Let A = (aij ) = [T ]E
E . Then for each , m,n ,
m
(e) X Y
(K(T )e , e ) = () a(t)(t) (Theorem 4.5.7)
|G| t=1
G
(e)
= d (A[|])
|G| G
(e) T
= d (A [|]) (Theorem 4.6.1(c))
|G| G
where A[|] = (a(i)(j) ) Cmm . In particular when m = n,
(e) (e) T
(K(T )e1 en , e1 en ) = d (A) = d (A ).
|G| G |G| G
We will use the above two theorems to establish some inequalities and iden-
tities. Notice that is irreducible if and only if is irreducible. Clearly
(e) = (e). So statements about dG (A) apply to dG (A).
Theorem 4.6.4. Let A, B Cmn . Then
|dG (AB )|2 dG (AA ) dG (BB ). (4.29)
Proof. Let ui = A(i) and vj = B(j) , where A(i) denotes the ith row of A. Then
(e)
(u , v ) = d ((ui , vj )) (Theorem 4.6.2)
|G| G
(e)
= d (AB ) ((u, v) := v u)
|G| G
Then apply Cauchy-Schwarz inequality
|(u , v )| (u , u )(v , v )
to have (4.29).
Theorem 4.6.5. Let A, B Cmm . Then
|dG (A)|2 (e) dG (AA ). (4.30)
Proof. In (4.29) pick B = Im and apply Theorem 4.6.1 to have (4.30).

Theorem 4.6.6. (Schur) If A Cmm is psd, then
(e) det A dG (A). (4.31)
Proof. Since A is psd, there is an upper triangular matrix L such that A = LL

(Problem 1.5 #9). By Theorem 4.6.1(b),
|dG (L)|2 = (e)2 | det L|2 = (e)2 det A.
Applying (4.30),
|dG (L)|2 (e)dG (LL ) = (e)dG (A).
From Theorem 4.6.6, if A 0, then
dG (A) 0, det A per A, det A h(A).
Theorem 4.6.7. (Fischer) If A Cmm is psd, then
det A det A[1, . . . p|1, . . . , p] det A[p + 1, . . . , m|p + 1, . . . , m]. (4.32)
Proof. This is indeed a special case of Theorem 4.6.4. Let G < Sm be the group
consisting of , where is a permutation on {1, . . . , p}, and is a permutation
on {p+1, . . . , m}. Let () = ()(). Then is a linear irreducible character
of G (check!) and dG (A) is the right side of (4.32).
See Horn and Johnsons Matrix Analysis p.478 for a matrix proof of Theorem
4.6.7.
Theorem 4.6.8. Let A Cmm . Then
1 X
h(A) = (e)dG (A).
|G|
I(G)
In particular, if A 0, then
(e) 1
h(A) d (A), h(A) per A
|G| G m!
Proof. Since A Cmm , there are u1 , . . . , um ; v1 , . . . , vm V where dim V = m

such that aij = (ui , vj ) (Problem 1.6 #3). Then
1 X X
(e)dG (A) = (u , v ) (by Theorem 4.6.2)
|G|
I(G) I(G)
X
= (T (G, )u , T (G, )v )
I(G)
X
= ( T (G, )u , v ) (by Theorem 4.2.3)
I(G)
m
Y
= (ui , vi ) (by Theorem 4.2.3(b))
i=1
= h(A).
P
When A 0, by Theorem 4.6.6 each summand dG (A) in 1
|G| I(G) (e)dG (A)
is nonnegative so we have
(e)
h(A) d (A)
|G| G
1
for all I(G). In particular with G = Sm and 1, we have h(A) m! per A.
We can express u = v in terms of the generalized matrix function when

u1 , . . . , um are linearly independent (so m dim V ).
Theorem 4.6.9.PLet u1 , . . . , um V be linearly independent. Then u = v if

m
and only if vi = j=1 aij uj , j = 1, . . . , m and dG (AA ) = dG (A) = (e).
Proof. Since u1 , . . . , um are linearly independent, by Theorem 4.3.2, u 6= 0.

() Suppose u = v (6= 0). From Theorem 4.3.3,
hu1 , . . . , um i = hv1 , . . . , vm i.
Pm
So there is A Cmm such that vi = j=1 aij uj , i = 1, . . . , m. Equip an inner
product on V so that {u1 , . . . , um } is an orthonormal set. Then
(ui , uj ) = ij , (vi , uj ) = aij , (vi , vj ) = (AA )ij .
Apply Theorem 4.6.2 we have

(e) (e)
(v , u ) = d ((vi , uj )) = d (A)
|G| G |G| G
(e) 2 (e)
(u , u ) = dG (Im ) =
|G| |G|
(e) (e)
(v , v ) = dG ((vi , vj )) = d (AA )
|G| |G| G
Then from u = v we have dG (AA ) = dG (A) = (e).

() From the assumption and the above argument,
(v , u ) = (u , u ) = (v , v ) (4.33)
so that
|(v , u )|2 = (u , u )(v , v )
with u 6= 0. By the equality case of Cauchy-Schwarz inequality, v = cu for
some constant. Then substitute into (4.33) to have c = 1.
Theorem 4.6.10. For all psd A, B Cmm ,
dG (A + B) dG (A) + dG (B).
In particular, if A B, then dG (A) dG (B).

Proof. Let E be an orthonormal basis of V with dim V = m. There are psd
S, T End V such that [S]E E
E = A and [T ]E = B. Notice that K(S + T )
K(S) + K(T ) (Problem 4.5 #3) and [T + S]E E = A + B. Apply Theorem 4.6.3
(switching to ) to have the desired result.
Suppose A B, i.e., A = B + C where C 0. The
dG (A) dG (B + C) dG (B) + dG (C) dG (B)
since dG (C) 0 by Theorem 4.6.5.

We now discuss the generalized Cauchy-Binet Theorem (compare Theorem
1.10.1).
Theorem 4.6.11. (Generalized
P Cauchy-Binet) Let A, B Cnn . For any
, = { : m,n : G () 6= 0}
(e) X
dG ((AB)[|]) = dG (A[|])dG (B[|])
|G|

Proof. Let E = {e1 , . . . , en } be an orthonormal basis of V . There are S, T

End V such that [S]E E
E = A and [T ]E = B. Notice that E = {e : m,n } is
m
an orthonormal basis of V with respect to the induced inner product. We
have
(e)
d ((AB)[|]
|G| G
= (K(ST )e , e ) (by Theorem 4.6.3)
= (K(S)K(T )e , e ) (K(ST ) = K(S)K(T ))
= (K(T )e , K(S )e ) (K(S) = K(S ))
X
= (K(T )e , e
)(e , K(S )e ) (by Theorem 1.3.3)
m,n
X
= (K(T )e , e )(e , K(S )e ) (by Theorem 4.2.4 m V =I(G) V (G))
m,n
X
= (K(S)e , e )(K(T )e , e ) (e 6= 0 , if m,n )

(e)2 X
= dG (A[|])dG (B[|]). (by Theorem 4.6.3)
|G|2

Then switch back to .
When G = Sm and = it reduces to Theorem 1.10.1.
Theorem 4.6.12. Let H < G < Sm and I(G). Suppose |H I(H).

Then for all psd A Cmm ,
1 1
d (A) d (A).
|H| H |G| G
Proof.
(e)2 X X
T (G, )T (H, ) = ()()P ()
|G||H|
G H
(e)2 X X
= ( 1 )()P ( ) ( = )
|G||H|
G H
(e) X
= ( )P ( ) (Theorem 2.4.3 since ( 1 ) = ( 1 ))
|G|
G
= T (G, ).
With respect to the induced inner product on m V where V is an m-dimensional

inner product space, T (G, ) is an orthogonal projection by Theorem 4.2.5(a).
So for any z m V
kzk kT (G, )zk.
Since T (G, )T (H, ) = T (G, ), apply the above inequality on T (H, )z,
kT (H, )zk kT (G, )T (H, )zk = kT (G, )zk.
Since A Cmm is psd and dim V = m, there are v1 , . . . , vm V such that

aij = (vi , vj ) (Problem 1.5 #10). Set z = v in the above inequality to have
(T (H, )v , T (H, )v ) (T (G, )v , T (G, )v ).
Then use Theorem 4.6.2 to have

1 1
dH (A) d (A).
|H| |G| G
The following famous conjecture is still open.
Permanent-on-top conjecture: If A Cmm is psd, then
dG (A) (e)per A.
Problems
1. Show that if A, B Cmm are psd, then det(A + B) det A + det B and
per (A + B) per A + per B.
Qm Pm
2. (Hadamard) Prove that if A Cmm , then | det A|2 i=1 j=1 |aij |2 .
Hint: det(AA ) h(AA ).
3. Prove that when is linear, Cauchy-Binet formula takes the form: for any

, ,
X 1
dG ((AB)[|]) = d (A[|])dG (B[|]).

|G | G

4. Suppose that H < G and assume that |H is also irreducible, where is

an irreducible character of G. Show that V (G) V (H).
5. If A Cmm is psd, then
per A per A[1, . . . , p|1, . . . , p] per A[p + 1, . . . , m|p + 1, . . . , m].

1. Follows from Theorem 4.6.10 with V = Cm , G = Sm , = ; V = Cm ,
G = Sm , 1.
2
2. By Theorem
Qm4.6.6,
Pm| det A| 2 = det(AA ) h(AA ) since AA is psd. But

h(AA ) = i=1 j=1 |aij | .
3. From Theorem 4.6.11, since (e) = 1,
1 X
dG ((AB)[|]) = dG (A[|])dG (B[|])
|G|

1 X X
= dG (A[|])dG (B[|]) ( = G)
|G|
G
Let A = (aij ) Cmm and let u1 , . . . , um ; v1 , . . . , vm V such that aij =

so that v = v(1) v(m) = 0.
(ui , vj ). If 6 G , then 6
By Theorem 4.6.2, if 6 G , then
1
d (A[|]) = (u(i) , v(j)

) = 0.
|G| G
So we have the desired result.

4.
5.
Chapter 5
Exterior spaces and

completely symmetric
spaces
5.1 Exterior spaces

Let dim V = n. When G = Sm with m n, and = , V (Sm ), denoted by
m V is called the mth exterior space or Grassmannian space. Denote by
v = v1 vm
the decomposable element v . The induced operator K(T ) is denoted by Cm (T ).

Since the alternating character is linear, the induced matrix K(A) is defined by
(4.25) and is denoted by Cm (A). It is called the mth compound of A.
Theorem 5.1.1. Let dim V = n m. For m V ,

=
(a) = Gm,n , = Qm,n , dim m V = n .
m
(b) Let E = {e1 , . . . , en } be a basis of V . Then E := {e

: Q
m,n } is a
basis of m V . If E is an orthonormal basis of V , then E0 := m!E is
an orthonormal basis of m V .
(c) For each Sm , v = ()v . When i 6= j and vi = vj , v = 0.
(d) v1 vm = 0 if and only if v1 , . . . , vm are linearly dependent.

Pm
(e) If ui = j=1 aij vj , i = 1, . . . , m, then u = det(aij )v .
Pm
(f) If u = v 6= 0, then ui = j=1 aij vj i = 1, . . . , m and det(aij ) = 1.
Pn P
(g) If vi = j=1 aij ej , i = 1, . . . , m, then v = Qm,n det A[1, . . . , m|]e
.
127
128CHAPTER 5. EXTERIOR SPACES AND COMPLETELY SYMMETRIC SPACES
Proof. The proofs of (a), (b), (c) are obtained in Chapter 4. Notice that (e) = 1
and Qm,n , G = {e}.
(d) By Theorem 4.3.2 it suffices to show

Pthe sufficiency. Let v1 , . . . , vm be
linearly dependent. Then for some j, vj = t6=j ct vt . So by (c)
X
v1 vm = ct v1 vj1 vt vj+1 vm = 0
t6=j
because each summand has repeated vs.
(e) Direct computation yields
m
X Xm
u = ( a1j vj ) ( amj vj )
j=1 j=1
X m
Y
= ( ai(i) )v
m,m i=1
X m
Y
= ( ai(i) )v ((d) and Dm,m = {(1, . . . , m) : Sm })
Dm,m i=1
m !
X Y
= ai(i) v
Sm i=1
m
!
X Y
= () ai(i) v
Sm i=1

= det(aij )v (Problem 1)
(f) By Theorem 4.3.3
hu1 , . . . , um i = hv1 , . . . , vm i.
Pm
Then ui = j=1 aij vj , i = 1, . . . , m. Use (e) to have det(aij ) = 1 (a special
case of Theorem 4.6.9)
5.1. EXTERIOR SPACES 129
(g) Similar to (e)

Xn Xn
v = ( a1j ej ) ( amj ej )
j=1 j=1
m
X Y
= ai(i) e

m,n i=1
m
X Y
= ai(i) e
((d) and Dm,n = { : Qm,n , Sm })
Dm,n i=1
X m
X Y
= ( ai(i) )e

Qm,n Sm i=1
m
!
X X Y
= () ai(i) e

Qm,n Sm i=1
X
= det A[1, . . . , m|]e

Qm,n
We now study the induced matrix Cm (A). From (4.26),

X Ym
1 1
K(A), = p () a(t),(t) = p dG (A[|]), (5.1)
|G ||G | G t=1
|G ||G |
Hence
for all , .
Cm (A), = det A[|],
for all , Qm,n .
For example, if n = 3 and k = 2, then

det A[1, 2|1, 2] det A[1, 2|1, 3] det A[1, 2|2, 3]
C2 (A) = det A[1, 3|1, 2] det A[1, 3|1, 3] det A[1, 3|2, 3] .
det A[2, 3|1, 2] det A[2, 3|1, 3] det A[2, 3|2, 3]
In general C1 (A) = A and Cn (A) = det A.

From Theorem 4.5.8 we have the following result.
Theorem 5.1.2. Let E = {e1 , . . . , en } be a basis of V and E := {e
:
Qm,n } and E = m!E , a basis of m V (in lexicographic order). Let T
0
End V and [T ]EE = A. For any , Qm,n ,

([Cm (T ))]E F
E ), = ([Cm (T ))]F ), = Cm (A), = det A[|].

The induced matrix Cm (A) is called the mth compound of A. Clearly
Cm (AB) = Cm (A)Cm (B).

Indeed it is the Cauchy-Binet determinant identity (Theorem 1.10.1) since

X
Cm (AB), = Cm (A), Cm (B), , , Qm,n .
Qm,n
By Theorem 5.1.2
X
det(AB)[|] = det A[|] det B[|].
Qm,n
We now list some basic properties of Cm (A).
Theorem 5.1.3. Let m n and A Cnn with eigenvalues 1 , . . . , n and

singular values s1 , . . . , sn . Then
(a) Cm (A ) = Cm (A) , Cm (A1 ) = Cm (A)1 if A is invertible.

r
(b) If rank A = r m, then rank Cm (A) = m . If r < m, then Cm (A) = 0.
(c) If A is upper triangular, so is Cm (A).

Qm Qm
(d) t=1 (t) are the eigenvalues and t=1 s(t) ( Qm,n ) are the singular
values of Cm (A).
P Qm P
(e) tr Cm (A) = Qm,n t=1 (t) = Qm,n det A[|].
(f) (Sylvester-Franke) det Cm (A) = (det A)(m1) .

n1
(g) If A is normal, Hermitian, pd, psd or unitary, so is Cm (A).
Compound matrix is a very powerful tool and the following results are two
applications.
Theorem Pn 5.1.4. Let A Cnn with eigenvalues |1 | |n |. Denote by

Ri = j=1 |aij |, i = 1, . . . , n, and let R[1] R[n] be an rearrangement of
R1 , . . . , Rn in nonincreasing order. Then
m
Y m
Y
|i | R[i] , 1 m n. (5.2)
t=1 t=1
Proof. Let Ax = 1 x where x is an eigenvector of A. Then

n
X
aij xj = 1 xi , i = 1, . . . , n.
j=1
Let 1 s n be an integer such that
|xs | = max |xi | > 0.

i=1
Then
n n n
X xj
X xj X

|1 | = asj a |asj | = Rs R[1] .
xs j=1 xs j=1
sj
j=1
Use Theorem 5.1.3(d) and apply the inequality on Cm (A) to have

m
Y
|i (A)| = |1 (Cm (A))| R[1] (Cm (A)).
i=1
For any Qm,n , use Theorem 5.1.2

X
R (Cm (A)) = | det A[|]|
Qm,n

X X m
Y

= () a(i)(i)

Qm,n Sm i=1
X m
Y
|a(i)(i) | (Dm,n = { : Qm,n , Sm })
Dm,n i=1
m
X Y
|a(i)(i) |
m,n i=1
m X
Y n m
Y m
Y
= |a(i)j | = R(i) R[i]
i=1 j=1 i=1 i=1
We can use (5.8) to have
| det A|
|n | Qn1 (5.3)
i=1 R[i]
since
n
Y
det A = i .
i=1
Theorem 5.1.5. Suppose that A Cnn is invertible with singular values

s1 sn > 0. Let A = QR be the unique QR decomposition where
positive diagonal entries of R. Let diag R = (a1 , . . . , an ). Let a[1] a[n]
be the rearrangement of a1 , . . . , an . Then
m
Y m
Y n
Y n
Y
a[t] st , m = 1, . . . , n 1, a[t] = st . (5.4)
t=1 t=1 t=1 t=1
Proof. We first establish a[1] s1 . Let A = QR be the QR decomposition of

A. Recall that s1 is the square root of the largest eigenvalue of the psd
A A = R Q QR = R R.
It is known that any diagonal entry of a Hermitian B is less than or equal to

the largest eigenvalue of B (by using Spectral Theorem for Hermitian matrices).
Suppose a[1] = ar for some 1 r n. Then
a2[1] = |ar |2 (R R)rr s21 . (5.5)
Hence a[1] s1 . and use Theorem 5.1.3 (c) and (g),
Cm (A) = Cm (Q)Cm (R)
is the QR decomposition of Cm (A). NowQthe diagonal entries of Cm (R) are

Q m m
i=1 a(i) Qm,n and the largest
Qm is i=1 a[i] By Theorem 5.1.3(d), the
largest singular value of Cm (A) is i=1 si . So by (5.5) we have
m
Y m
Y
a[i] si , i = 1, . . . , n 1.
i=1 i=1
The equality is obtained via the determinant equation
det A = det Q det R
so that
n
Y n
Y
si = | det A| = | det R| = a[i] .
i=1 i=1
Remark: The converse of Theorem 5.1.5 is true, i.e., given positive numbers
a1 , . . . , an and s1 sn > 0, if (5.4) is true, then there exists A Cnn
with singular values ss and diag R = (a1 , . . . , an ) where A = QR is the QR
decomposition of A. This result is known as Kostants convexity theorem [11].
Indeed Kostants result is in a broader context known as real semisimple Lie
groups.
For A Cnn , t C, we write
(1)
Cm (In + tA) = I( n ) + tDm (A) + t2 Dm
(2)
(A) + + tm Dm
(m)
(A).
m
(r)
The matrix Dm (A) is called the rth derivative of Cm (A). In particular
(1)
Dm (A) is the directional derivative of Cm () at the identity In in the direc-
tion A. It is viewed as the linearization of Cm (A)
(r)
Theorem 5.1.6. If A Cnn is upper triangular, then Dm (A) is upper tri-
angular.
Proof. If A is upper triangular, so is In + tA. Thus Cm (In + tA) is upper

triangular.
(r) (r)
Theorem 5.1.7. (a) Dm (SAS 1 ) = Cm (S)Dm (A)Cm (S)1 for all 1 r
m. So, if A and B are similar, then Cm (A) and Cm (B) are similar.
(b) Let A Cnn with eigenvalues 1 , . . . , n . Then the eigenvalues of

(r)
Dm (A) are
r
X Y
(i) , Qm,n .
Qr,m i=1
(1)
In particular (1) + + (m) , Qm,n , are the eigenvalues of Dm (A).
(r) (r)
(c) Dm (A ) = Dm (A) .
Proof. (a)
Cm (In + tSAS 1 )
= Cm (S(In + tA)S 1 )
= Cm (S)Cm (In + tA)Cm (S 1 )
= Cm (S)Cm (In + tA)Cm (S)1
(1)
= Cm (S)(I( n ) + tDm (A) + t2 Dm
(2)
(A) + + tm Dm
(m)
(A))Cm (S)1
m
(1)
= I( n ) + tCm (S)Dm (A)Cm (S)1 + t2 Cm (S)Dm
(2)
(A)Cm (S)1
m
+ + tm Cm (S)Dm
(m)
(A)Cm (S)1
On the other hand
Cm (In +tSAS 1 ) = I( n ) +tDm

(1)
(SAS 1 )+t2 Dm
(2)
(SAS 1 )+ +tm Dm
(m)
(SAS 1 ).
m
Comparing coefficients of tr of both equations yields the desired result.

(b) By Schur triangularization theorem, there is U U(n) such that U AU =
T where T is upper triangular and diag T = (1 , . . . , n ). By (a)
(r)
Dm (A) = Cm (U )Dm
(r)
(T )Cm (U )
(r) (r)
and hence Dm (A)) and Dm (T )) have the same eigenvalues. From Theorem
(r)
5.1.6 Dm (T ) is upper triangular so the diagonal entries are the eigenvalues of
A. Now
Cm (I + tT ),
= det(In + tT )[|]
Ym
= (1 + t(i) )
i=1
2
X Y
= 1 + t((1) + + (m) ) + t2 (i)
Q2,m i=1
r
X Y
+ + tr (i) +
Qr,m i=1
(r) P Qr
for all Qm,n . So the eigenvalues of Dm (A) are Qr,m i=1 (i) ,
Qm,n .
(c) For any t R,
(1)
I( n ) + tDm (A ) + t2 Dm
(2)
(A ) + + tm Dm
(m)
(A )
m
= Cm (I + tA )
= Cm (I + tA)
(1)
= I( n ) + tDm (A) + t2 Dm(2)
(A) + + tm Dm(m)
(A)
m
(r) (r)
Hence Dm (A ) = Dm (A) for all r as desired.
(1)
Let A Cnn and set m (A) := Dm (A). It is called the mth additive
compound of A.
Theorem 5.1.8. m : Cnn C( n )( n ) is linear, i.e., m (A+B) = m (A)+
m m
m (B) for all A, B Cnn and m (cA) = cm (A).
Proof. By Theorem 5.1.3(a)
Cm ((In + tA)(I + tB)) = Cm (In + tA)Cm (In + tB).
Now
Cm (In + tA)Cm (In + tB) = (I( n ) + tm (A) + o(t2 ))(I( n ) + tm (B) + o(t2 ))
m m
= I( n ) + t(m (A) + m (B)) + o(t2 ),

m
and
Cm ((In +tA)(In +tB)) = Cm (In +t(A+B)+t2 AB) = I( n ) +t(m (A+B))+o(t2 ).

m
Then compare coefficients of t on both sides. The equality m (cA) = cm (A)

is easily deduced as well.
The following problem is still open
Marcusde Oliveira Conjecture

Let A, B Cnn be normal matrices with eigenvalues 1 , . . . , n and 1 , . . . , n .
The Marcusde Oliveira Conjecture claims that
n
Y
1 1
{det(V AV + U BU ) : V, U U(n)} conv { (i + (i) ) : Sn },
i=1
where conv S denotes the convex hull of the set S C.

Since
det(V AV 1 + U BU 1 ) = det(V (A + V 1 U BU 1 V )V 1 )
= det V det(A + V 1 U BU 1 V ) det V
= det(A + V 1 U BU 1 V ),
by using spectral theorem for normal matrices, Marcusde Oliveira Conjecture

can be restated as
{det(diag (1 , . . . , n ) + U diag (1 , . . . , n )U 1 ) : U U(n)}

Yn
conv { (i + (i) ) : Sn },
i=1
where conv S denotes the convex hull of the set S C.

When A and B are Hermitian, the inclusion becomes an equality, a result
of Fielder [2]. See the generalization of Fielders result in the context of real
semisimple Lie algebras [28, 29, 30].
For A, B Cnn , the Cauchy-Binet formula yields [21]

In I
det(A + B) = det((A In ) ) = Cn (A In )Cn n .
B B
Thus

In In
det(A+U BU 1 ) = Cn (A In )Cn = Cn (A In )Cn ((U U ) U 1 ).
U BU 1 B
Problems
1. Show that for m V (m n = dim V ), = Dm,n .
2. Let T End V and dim V = n. Prove that for all v1 , . . . , vn V ,
Cn (T )v1 vn = (det T )v1 vn .
3. Suppose V (G) 6= 0 and v1 vm = 0 whenever v1 , . . . , vm are linearly
dependent. Prove that V (G) = m V , i.e., G = Sm and = .
Pk
4. (Cartan Lemma) Let e1 , . . . , ek be linear independent and i=1 ei vi =
0. Prove that there is a symmetric matrix A Ckk such that vi =
Pk
i=1 aij ej , i = 1, . . . , k. Hint: Extend e1 , . . . , ek to a basis e1 , . . . , en
of
PnV . Notice that P {ei ej : 1 i < j n} is a basis of 2 V and
i,j=1 aij ei ej = i<j (aij aji )ei ej .
qP
n 2
5. Let A Cnn and Ei = j=1 |aij | and use similar notations as in
Theorem 5.1.4. Prove that
Ym Y m
n
|i | E[i] , m = 1, . . . , n
i=1
m i=1
and
| det A|
|n | Qn1 .
n i=1 E[i]
Hint: Use Problem 4.6 #2.
6. (Weyl) Let A Cnn . Let 1 , . . . , n with the ordering |1 | | n |

and s1 sn be the singular values of A, respectively. Prove that
m
Y m
Y n
Y n
Y
|t | st , m = 1, . . . , n 1, |t | = st . (5.6)
t=1 t=1 t=1 t=1
(Hint: Establish |1 | s1 and use Theorem 5.1.3(d). Remark: The

converse is true, i.e., given 1 , . . . , n in nonincreasing moduli and s1
sn 0, if (5.6) is true, then there exists A Cnn with eigenvalues
s and singular values ss. It is due to A. Horn [7] and see a different
proof of T.Y. Tam [27]).
7. Let A Cnn with distinct eigenvalues 1 , . . . , n . Define
Sep(A) = min |i j |.
i6=j
Set L := A In In A. Prove
!1/2
|tr Cn2 n (L)|
Sep(A) Qn2 n2 .
i=1 R[i] (L)
Hint: The eigenvalues of L are i j i, j = 1, . . . , n, in which n2 n are

nonzero so Cn2 n (L) has only one nonzero eigenvalue
Y n2 n Y
(i j ) = (1) 2 (i j )2 = tr Cn2 n (L).
i6=j i<j
Then apply (5.3).

5.2. DECOMPOSABLE ELEMENTS IN THE EXTERIOR SPACES 137
8. Let A Cnn be Hermitian with distinct eigenvalues and set L := A

In In A. Prove
!1/2
| det L[|]|
Sep(A) Qn2 n2 ,
i=1 R[i] (L)
where Qn2 n,n2 . Hint: |1 | maxi=1,...,n |aii |.
9. Let 1 , . . . , n be the eigenvalue of A Cnn . Prove that (i j )2 ,

(1)
1 i < j n, are the eigenvalues of M := D2 (A2 ) 2C2 (A). When
the eigenvalues of A are distinct, use M to give a nonzero lower bound of
Sep(A).
Solutions
1.
2.
3.
4.
5.
6.
7.
5.2 Decomposable elements in the exterior spaces

We will give some characterization of decomposable elements in the exterior
space m V . We know from Theorem 5.1.1(d) that v = v1 vm is nonzero
if and only if v1 , . . . , vm are linearly independent. PmMoreover from Theorem
5.1.1(f) if v 6= 0, then u = v if and only if ui = j=1 cij vj , i = 1, . . . , m and
det(cij ) = 1. So the two m-dimensional subspaces hu1 , . . . , um i and hv1 , . . . , vm i
of V are the same if and only if u = cv 6= 0. Thus nonzero decomposable
v are in one-one correspondence (up to a nonzero scalar multiple) with m-
dimensional subspaces of V . The space of m-dimensional subspaces of V is
called the mth Grassmannian of V .
Suppose that E = {e1 , . . . , en } is a basis of V and {v1 , . . . , vm } is a basis for
the m-dimensional subspace W of V . Since vs are linear combinations of es,
from Theorem 5.1.1(g) X
v = a e
.
Qm,n
n
So up to a nonzero scalar multiple, those m scalars a are uniquely determined
by W . In other words, if we choose another basis {u1 , . . . , um } of W so that
X
u = b e
,
Qm,n
then
a = cb , for all Qm,n
n
for some constant c 6= 0. The m scalars a are called the Plucker coordi-
nates of v . The concept was introduced nby
Julius Plucker in the 19th century
for studying geometry. Clearly not all m scalars are Plucker coordinates for
some v since not all vectors in m V are decomposable. We are going to give
a necessary and sufficient condition for Plucker coordinates.
By see Theorem 5.1.1(c),
X X
a e
= v = ()v =

()a e

Qm,n Qm,n
so that
a = ()a .
So it is a necessary condition and motivates the following definition.
Let p : m,n C be a function satisfying
p() = ()p(), Sm , m,n . (5.7)

P
For any element z = Qm,n c e m
V , the coordinates c can be viewed as
a function c(). This function can be extended to m,n such that (5.7) remains
true. Conversely, any pPsatisfying (5.7) gives a one-one correspondence between
elements of m V and Qm,n p()e .
For any , m,n and p satisfying (5.7), define P (, ) Cmm such that
P (, )i,j = p([i, j : ]), 1 i, j m,
where [i, j : ] m,n is obtained from by replacing (i) by (j), i.e.,
[i, j : ] = ((1), . . . , (i 1), (j), (i + 1), . . . , (m)) m,n .
The following result gives the relationship between decomposability and subde-
terminants.
Theorem 5.2.1. Let 1 m n. If p satisfies (5.7) and is not identically zero,
then the following are equivalent.
P
1. z = Qm,n p()e is decomposable.
2. There is A Cmn such that p() = det A[1, . . . , m|], Qm,n .

3. det P (, ) = p()m1 p(), , Qm,n .
Proof.
P
(a) (b). Suppose z = Qm,n p()e is decomposable, i.e., z = v1
Pn
vm . Write vi = j=1 aij ej , i = 1, . . . , m, where E = {e1 , . . . , en } is a basis
of V . By Theorem 5.1.1(g),
X
z = v = det A[1, . . . , m|]e

Qm,n
Comparing coefficients yields p() = det A[1, . . . , m|], Qm,n . P

n
(b) (a). Suppose that (b) is true. Use A to define vi = j=1 aij ej ,
i = 1, . . . , m. So
X X
z= p()e
= det A[1, . . . , m|]e
= v1 vm
Qm,n Qm,n
is decomposable.
(b) (c). For any A Cmn and , Qm,n , let
S := A[1, . . . , m|] = (A(1) . . . A(m) ) Cmm
where A(i) denotes the ith column of S, i = 1, . . . , m. Similarly let
T := A[1, . . . , m|] = (A(1) . . . A(m) ) Cmm .
If we define X Cmm by
xij = det(A(1) . . . A(i1) A(j) A(i+1) . . . A(m) ), 1 i, j m,
then by Cramer rule
SX = (det S)T.
Taking determinants on both sides
det S det X = det T (det S)m .
If S is invertible, then
det X = det T (det S)m1 . (5.8)
If S is not invertible, then replace S by S + tIm . Since det B is a polynomial
function of its entries bij , we have (5.8) by continuity argument.
Suppose that (b) is true. Then
det S = det A[1, . . . , m|] = p(),
det T = det A[1, . . . , m|] = p()
and
xij = det(A(1) . . . A(i1) A(j) A(i+1) . . . A(m) )
= det A[1, . . . , m|(1), . . . , (i 1), (j), (i + 1), . . . , (m)]
= p((1), . . . , (i 1), (j), (i + 1), . . . , (m))
= p([i, j : ]).
Then X = (xij ) = P (, ) and (5.8) becomes
det P (, ) = p()m1 p().
Suppose (c) is true. Since p is not a zero polynomial, there is some Qm,n
such that p() 6= 0. Define A Cmn where
1
aij = p() m 1 p((1), . . . , (i1), (j), (i+1), . . . , (m)), 1 i m, 1 j n.
Then for any Qm,n ,
det A[1, . . . , m|]

= det(ai(j) )
1
= det(p() m 1 p((1), . . . , (i 1), (j), (i + 1), . . . , (m))
1
= ((p() m 1 )m det(p([i, j : ])1i,jm
= p()1m det P (, )
= (p()1m p()m1 p() by (c)
= p().
There are many equivalent conditions for decomposability, for example
P (, )P (, ) = p()P (, ), , , m,n
is also necessary and sufficient condition (but we will not discuss it in detail).
We now use Theorem 5.2.1 to deduce some results.
Let Dm,n . Denote by Im denotes the set of the components of . We
say that , Dm,n differ by k components if |Im Im | = m k. For
example = (1, 3, 5) and = (3, 5, 6) differ by one component. We say that
1 , . . . , k Qm,n form a chain if all consecutive sequences differ by one com-
ponent. The following is an important necessary condition for decomposability.
P
Theorem 5.2.2. Let z = Qm,n p()e k
V (with dim V = n) be de-
composable. If p() 6= 0, p() 6= 0 and and differ by k components, then
there are k 1 different sequences 1 , . . . , k1 Qm,n such that p(i ) 6= 0,
i = 1, . . . , k 1, and , 1 , . . . , k1 , form a chain. In other words, any two
nonzero coordinates of a decomposable v are connected by a nonzero chain.
Proof. When k = 1, the result is trivial. Suppose k > 1. Use (5.7) to extend
the domain of p to m,n . By Theorem 5.2.1
det P (, ) = p()m1 p() 6= 0
so one term in the expansion of the left side must be nonzero, i.e., there is
Sm such that
Ym
p([i, (i) : ]) 6= 0.
i=1
Since and differ by k > 1 components, there is i0 (1 i0 m) such that

(i0 ) 6 Im . From p([i0 , (i0 ) : ]) 6= 0 we know that ((i0 )) 6 Im and
[i0 , (i0 ) : ] Dm,n . So there is Sm such that [i0 , (i0 ) : ] = 1
Qm,n . By (5.7),
p(1 ) = ()p([i0 , (i0 ) : ]) 6= 0.
Clearly and 1 differ by one component and 1 and differ by k 1 compo-
nents. Repeat this process for 1 and to have 2 , . . . , k1 .
Example:
PTheorem 5.2.2 can be used to determine indecomposable elements,
i.e., if z = Qm,n p()e has two nonzero unchained coordinates, then z is
indecomposable. For example z = e1 e2 + e3 e4 is indecomposable since
p(1, 2) and p(3, 4) are nonzero but not chained.
Theorem 5.2.3. Let dim V = n and 1 m n. Then all elements in m V is
decomposable if and only if m = 1, m = n 1 or m = n.
Proof. Clearly if m = 1 or m = n, all elements of P m V are decomposable.
So we only need to consider m = n 1. Suppose z = Qm,n e m
V . By
(5.7) extend the domain of p to m,n and we are going to show that p satisfies
Theorem 5.2.1(c).
For any , Qm,n , if = , then P (, ) = p()Im . Hence (c) is valid.
If 6= , then from m = n 1, and differ by one component. So has
a unique component (t) 6 Im and there is Sm such that (j) = (j)
unless j = t. So
p([t, t : ]) = p((1), . . . , (t 1), (t), (t + 1), . . . , (m))
= p((1), . . . , (t 1), (t), (t + 1), . . . , (m))
= p().
When j 6= t,
p([i, j : ]) = p([i, j : ]) = p()ij .
So the (m1)(m1) principal submatrix P (, )(t|t) of P (, ) is p()Im1
and the tth diagonal entry of P (, ) is p(). So
det P (, ) = p()m1 p() = ()p()m1 p().
By the property of determinant,
det P (, ) = det p([i, (j) : ]) = () det p([i, j : ]) = () det P (, ).
Hence det P (, ) = p()m1 p(), i.e., Theorem 5.2.1(c) is satisfied. Hence all
elements of n1 V are decomposable.
Conversely if 2 m n 2, let
= (1, . . . , m), = (1, . . . , m 2, m + 1, m + 2).
Then consider z = e
+ e . Notice that and differ by two components and
are linked by a nonzero chain. By Theorem 5.2.2, z is indecomposable. So not
all elements of m V are decomposable.
P
Theorem 5.2.4. Let v1 vm = Qm,n p()e 6= 0. If p() 6= 0, let
Pn
ui = j=1 p([i, j])ej , i = 1, . . . , m. Then u1 um = p()m1 v1 vm .
Thus hu1 , . . . , vm i = hv1 , . . . , vm i. Here
[i, j] := ((1), . . . , (i 1), j, (i + 1), . . . , (m)) m,n .
Proof. Let aij := p([i, j]) = p((1), . . . , (i 1), j, (i + 1), . . . , (m)). By

Theorem 5.1.1, X
u = det A[1, . . . , m|]e .
Qm,n
By Theorem 5.2.1
det A[1, . . . , m|] = det(ai(j) = det(p([i, j : ]) = p()m1 p().
So u = p()m1 v . Then by p() 6= 0, v 6= 0 to have hu1 , . . . , , um i =

hv1 , . . . , vm i.
By Theorem 5.2.4, it is easy to find a basis of the subspace that corresponds
to a given decomposable element. For example, from Theorem 5.2.3,
z = e1 e2 e3 + 2e1 e2 e4 e1 e3 e4 3e2 e3 e4 3 V
is decomposable with dim V = 4. It is not hard to find u1 , u2 , u3 V such that

z = u1 u2 u3 : by Theorem 5.2.4, let = (1, 2, 3) so that p() = 1. Then
u1 = p(1, 2, 3)e1 + p(2, 2, 3)e2 + p(3, 2, 3)e3 + p(4, 2, 3)e4 = e1 3e4
because p(2, 2, 3) = p(3, 2, 3) = 0 and p(4, 2, 3) = p(2, 3, 4) by (5.7). Similarly

u2 = e2 + e4 , u3 = e3 + 2e4 . So
z = (e1 3e4 ) (e2 + e4 ) (e3 + 2e4 ) = 3.
Problems
P
1. Let v1 vm = Wm,n p()e
6= 0. Prove that for any m,n ,
n
X
ui = p([i, j])ej hv1 , . . . , vm i, i = 1, . . . , m.
j=1
2. It is known that z = 3e1 e2 +5e1 e3 +5e1 e4 +e2 e3 2e2 e4 +5e3 e4

is decomposable, find u1 , u2 V such that z = u1 u2 .
3. Let p : m,n C not identically zero. Suppose that p satisfies (5.7).
Prove that the following are equivalent.
(a) P (, )P (, ) = p()P (, ), , , m,n .
5.3. COMPLETELY SYMMETRIC SPACES 143
(b) P (, )P (, ) = p()p()Im , , Gmn.

Pm
(c) t=1 p([s, t : ])p([t, s : ]) = p()p(), , m,n , s = 1, . . . , m.
Pm+1 k1
(d) k=1 (1) p((i) : (k))p((k)) = 0, m,n , m+1,n ,
where
(i) : (k) = ((1), . . . , (i 1), (i + 1), . . . , (m), (k)) m,n

(k) = ((1), . . . , (k 1), (k + 1), . . . , (m + 1)) m,n .
5.3 Completely symmetric spaces

We denote V (G) by V1 (Sm ) = m V when G = Sm and 1. The space m V
is called a completely symmetric spaces. The decomposable element v is
written as v = v1 vm .
Let A Cmm , the permanent
m
X X
per A = at (t)
Sm t=1
plays an important role for the completely symmetric space as the determinant
to the exterior space.
Theorem 5.3.1. Let dim V = n m. Denote by () = |G | (G = Sm ). For

m V ,

=
(a) = = Gm,n and dim m V = n+m1 .
m
(b) If E = {e1 , . . . , en } is a basis of V , then E := {e : q Gm,n } is a basis of

m!
m V . If E is an orthonormal basis of V , then E0 := { () e : Gm,n }
m
is an orthonormal basis of V .
(c) For each Sm , v = v .
(d) v1 vm = 0 if and only if some vi = 0.
(e) If u = v 6= 0, then
Qm there is Sm and di 6= 0 such that ui = di v(i) ,
i = 1, . . . , m and i=1 di = 1.
Pn
(f) If vi = j=1 aij ej , i = 1, . . . , m, then
X 1
v = per A[1, . . . , m|]e .
()
Gm,n
Proof. (a), (b), (c) are clear from Chapter 4.

(d) It suffices to prove the necessary part and we use contrapositive. Equip
V with an inner product. Suppose vt 6= 0 for all t = 1, . . . , m. The union
m
i=1 vi of the finite hyperplanes vi is not dense in V . So there is w W such
that
Ym
(w, vt ) 6= 0.
t=1
On the other hand, for any w V
m
Y
(w w, v ) = (T (G, 1)w w, v ) = (w w, v1 vm ) = (w, vt ).
t=1
Qm
So v = 0 would imply t=1 (w, vt ) = 0 for all w V . Thus v 6= 0.
(e) If u = v 6= 0, then for any w V ,
m
Y m
Y
(w, ut ) = (w w, u ) = (w w, v ) = (w, vt ). (5.9)
t=1 t=1
Decompose ut =Qxt + yt such that xt hv1 i, yt v1 , t = 1, . . . , m. For any

m
z v1 , we have t=1 (z, vt ) = 0. By (5.9)
m
Y m
Y m
Y
0= (z, vt ) = (z, ut ) = (z, yt ), for all z v1 .
t=1 t=1 t=1
Applying the claim in the proof of (d) on the space v1 to have yi = 0 for some
i, i.e., ui = xi hv1 i. So there is di 6= 0 such that ui = di v1 . Substitute it into
(5.9)
m
Y Y
0 = (w, v1 )( (w, vt ) di (w, ut )). (5.10)
i=2 t6=i
Since v1 6= 0 and w V is arbitrary, we have

m
Y Y
(w, vt ) = di (w, ut ), for all w v1
i=2 t6=i
Then apply continuity argument, so that the above equality is valid for w V .
Then use induction.
(f)
n
X Xn

v = ( a1j ej ) ( amj ej )
j=1 j=1
X m
Y
= ai(i) e
m,n i=1
X m
1 X Y
= ai(i) e (Theorem 4.4.3)
() i=1
Gm,n Sm
X 1
= per A[1, . . . , m|]e .
()
Gm,n
5.3. COMPLETELY SYMMETRIC SPACES 145
We denoted by Pm (T ) for K(T ) if T End V and Pm (A) for K(A) if

A Cnn .
Theorem 5.3.2. Let T End V with dim V = n. Let 1 , . . . , n be the
eigenvalues and s1 , . . . , sn be the singular values of A. Let A = [T ]E
E Cnn
where E is a basis of V . Then
(a) Let E0 be the basis as in Theorem 5.3.1. Then
per A[|]
E0
[Pm (T )]E0 = Pm (A), = p .
, ()()
(b) Pm (A ) = Pm (A) , Pm (A1 ) = Pm (A)1 if A is invertible, Pm (AB) =

Pm (A)Pm (B).

(c) If rank A = r, then rank Pm (A) = r+m1
m .
(d) If A is upper triangular, so is Pm (A).
Qm Qm
(e) t=1 (t) are the eigenvalues and t=1 s(t) are the singular values,
Gm,n , of Pm (A)
P Qm P 1
(f) tr Pm (A) = Gm,n t=1 (t) = Gm,n () per A[|].
(g) det Pm (A) = (det A)( n ) .

n+m1
(h) If A is normal, Hermitian, pd, psd or unitary, so is Pm (A).
Problems
1. Prove the Cauchy-Binet formula for the permanent: Let A, B Cnn .
For any , Gm,n ,
X 1
per ((AB)[|]) = per (A[|])per (B[|]).
()
Gm,n
2. Prove Laplace expansion for permanent: for any A Cnn , Qm,n ,

X
per A = per A[|]per A(|).
Qm,n
3. Let G < Sm . Check which properties of Theorem 5.3.1 remain true for
V1 (G).
4. Prove that the set {v v : v V } generates V m = V1 (Sm ).
5. Let A(t) Cnn be a differentiable matrix function. Prove that
Xn
d daij (t)
per A(t) = per A(i|j) .
dt i,j=1
dt
Chapter 6
Research topics
6.1 Orthonormal decomposable basis

The material in this section is from Holmes [3].
Let V be a finite-dimensional complex inner product space and assume n :=
dim V 2 (to avoid trivialities). Let G < Sm . Fix an orthonormal basis
{e1 , . . . , en } of V . Given m,n and I(G), the inner product on V
induces an inner product on m V . If the subspace W of m V has a basis
consisting of mutually orthogonal standard symmetrized tensors, we will say
that W has an o-basis.
Given , set G := { G | = } G. We have
(1) X (1) X
(e , e ) = ( 1 ) = ( 1 ),
|G| |G|
G G
the first equality from [1, p. 339] and the second from the observations that
G 1 = G and ( 1 ) = ( 1 1 ).
We have dim V = (1)(, 1)G .
The group G is 2-transitive if, with respect to the componentwise action,
it is transitive on the set of pairs (i, j), with i, j = 1, . . . , m, i 6= j. Note that if
G is 2-transitive, then for any i = 1, . . . , m, the subgroup { G | (i) = i} of
G is transitive on the set {1, . . . , i, . . . , m}.
Theorem 6.1.1. Assume m 3. If G is 2-transitive, then m V does not have

an o-basis.
Proof. By the remarks above, it is enough to show that V does not have an
o-basis of some I(G), .
Let H = { G | (m) = m} < G and denote by the induced character
(1H )G , so that () = |{i | (i) = i}| for G (see [4, p. 68]).
Let G H and for i = 1, . . . , m, set Ri = { H | (i) = i}. Clearly,
Ri = if i {m, 1 (m)}. Assume i / {m, 1 (m)}. Since H acts transitively
147
148 CHAPTER 6. RESEARCH TOPICS
on 1, . . . , m 1, there exists some Ri . Then Ri = Hi , where Hi := {

H | (i) = i}. Now [H : Hi ] equals the number of elements in the orbit of i
under the action of H, so [H : Hi ] = m 1. Therefore, |Ri | = |Hi | = |Hi | =
|H|/[H : Hi ] = |H|/(m 1). We obtain the formula
X n
X X m2
() = |Ri | = |Ri | = |H|.
i=1
m1
H i6=m,1 (m)
Since (, 1)G = (1, 1)H = 1 by Frobenius reciprocity, 1 is a constituent of ,

whence := 1 is a character of G. Moreover, the 2-transitivity of G implies
that (, )G = 2 (see [4, p. 68]). Hence, (, )G = 1, so that is irreducible.
Let = (1, . . . , 1, 2) and note that G = H. Let and be represen-
tatives of distinct right cosets of H in G. Then := 1 G H, so the
discussion above shows that

(1) X (1) n 2
(e , e ) = ( 1 ) = |H| |H| < 0.
|G| |G| m 1
H
It follows that distinct standard symmetrized tensors in V are not orthogonal.

On the other hand,
dim V = (1)(, 1)H = (m 1)[(, 1)H 1],
and since (, 1)H = (, )G = 2 by Frobenius reciprocity, dim V = n 1 > 1.

Therefore, V does not have an o-basis. This completes the proof.
Corollary 6.1.2. If G = Sm (m 3) or G = Am (m 4), then m V does not
have an o-basis.
2-transitive groups have been studied extensively (see [3, Chapter XII], for
example).
Problems
1. Show that if G is abelian, then m V has an o-basis.
2. Give a proof of Corollary 6.1.2 by showing that (i) Sm is 2-transitive for
all m and (ii) the alternating group Am is 2-transitive if m 4.
3. Show that if G is the dihedral group Dm Sn , then m V has an o-basis
if and only if m is a power of 2.
Bibliography
[1] A.C. Aitken, The normal form of compound and induced matrices, Proc.
London Math. Soc. 38 (1934), 354376.
[2] M. Fiedler, Bounds for the determinant of the sum of hermitian matrices,
Proc. Amer. Math. Soc., 30 (1971), 2731.
[3] R.R. Holmes, Orthogonal bases of symmetrized tensor spaces, Linear and
Multilinear Algebra 39 (1995), 241243.
[4] R.R. Holmes and T.Y. Tam, Symmetry classes of tensors associated with
certain groups, Linear and Multilinear Algebra 32 (1992), 2131.
[5] R.R. Holmes, C.K. Li, T.Y. Tam, Numerical range of the derivation of an
induced operator. J. Aust. Math. Soc. 82 (2007), 325-3-43.
[6] C. Gamas, Conditions for a symmetrized decomposable tensor to be zero,
Linear Algebra Appl., 108 (1988), 83-119.
[7] A. Horn, On the eigenvalues of a matrix with prescribed singular values,
Proc. Amer. Math. Soc., 5 (1954) 47.
[8] I.M. Isaacs, Character Theory of Finite Groups, Academic Press, New York,
1976.
[9] G. James and A. Kerber, The Representation Theory of the Symmetric
Group, Encyclopedia of Mathematics and Its Applications, Volume 16,
Addison-Wesley Publishing Company, 1981.
[10] G. James and M. Liebeck, Representations and Characters of Groups, Cam-
bridge Mathematical Textbooks, Cambridge University Press, 1993.
[11] B. Kostant, On convexity, the Weyl group and Iwasawa decomposition,
Ann. Sci. Ecole Norm. Sup. (4), 6 (1973) 413460.
[12] C.K. Li, and T.Y. Tam, Operator properties of T and K(T ), Linear Algebra
Appl. 401 (2005), 173191.
[13] D.E. Littlewood, On induced and compound matrices, Proc. London Math.
Soc. 40 (1936), 370380.
149
150 BIBLIOGRAPHY
[14] M. Marcus, Finite Dimensional Multilinear Algebra, Part I, Marcel Dekker,

New York, 1973.
[15] M. Marcus, Finite Dimensional Multilinear Algebra, Part II, Marcel

Dekker, New York, 1975.
[16] R. Merris, Nonzero decomposable symmetrized tensors, Linear and Multi-

linear Algebra 5 (1977), 287292.
[17] R. Merris, On vanishing decomposable symmetrized tensors, Linear and

Multilinear Algebra, 5 (1977/78), 79-86.
[18] R. Merris, The structure of higher degree symmetry classes of tensors II,
Linear and Multilinear Algebra, 6 (1978), 171178.
[19] R. Merris, Recent advances in symmetry classes of tensors, Linear and

Multilinear Algebra , 7 (1979) 317-328.
[20] R. Merris, Multilinear Algebra, Gordon and Breach Science Publishers, Am-
sterdam, 1997.
[21] U. Prells, M.I. Friswell, S.D. Garvey, Use of geometric algebra: Compound
matrices and the determinant of the sum of two matrices, Proc. R. Soc.
Lond. A, 459 (2030), 273285.
[22] J.A. Dias da Silva, Conditions for equality of decomposable symmetric

tensors, Linear Algebra and its Appl., 24 (1979), 8592.
[23] J.A. Dias da Silva, Colorings and equality of tensors, Linear Algebra Appl.,
342 (2002) 7991.
[24] J.A. Dias da Silva, J. A. and M.M. Torres, On the orthogonal dimension
of orbital sets, Linear Algebra Appl., 401 (2005), 77107.
[25] , J.A. Dias da Silva and H.F. da Cruz, Equality of immanantal decompos-
able tensors, Linear Algebra Appl., 401 (2005), 2946.
[26] J.P. Serre, Linear Representations of Finite Groups, Springer, New York,
1977.
[27] T.Y. Tam, A. Horns result on matrices with prescribed singular values and
eigenvalues, Electronic Linear Algebra, 21 (2010) 2527.
[28] T.Y. Tam and M.C. Thompson, Determinant and Pfaffian of sum of skew
symmetric matrices, Linear Algebra Appl., 433 (2010) 412423.
[29] T.Y. Tam and M.C. Thompson, Determinants of sum of orbits under com-
pact Lie group, to appear in Linear Algebra Appl.
[30] T.Y. Tam and M.C. Thompson, Determinants of sums of two real matrices
and their extensions, submitted for publication.
BIBLIOGRAPHY 151
[31] J.H. van Lint, Notes on Egoritsjevs proof of the van der Waerden conjec-
ture. Linear Algebra Appl., 39 (1981), 18.
[32] J.H. van Lint, The van der Waerden conjecture: two proofs in one year.
Math. Intelligencer 4 (1982), 7277.
[33] B.Y. Wang, Foundation of Multilinear Algebra, Beijing Normal University

Press (in Chinese), 1985.
[34] B.Y. Wang, Wang and M.P. Gong, A high symmetry class of tensors with
an orthogonal basis of decomposable symmetrized tensors, Linear and Mul-
tilinear Algebra, 30 (1991), 6164.
[35] B.Y. Wang, Wang and M.P. Gong, The subspaces and orthonormal bases
of symmetry classess of tensor, Linear and Multilinear Algebra, 30 (1991),
195204.
[36] H. Weyl, Inequalities between the two kinds of eigenvalues of a linear trans-
formation, Proc. Nat. Acad. Sci. U. S. A., 35 (1949) 408411.

Tin-Yau Tam Multilinear Algebra (Lecture Notes)

Încărcat de

Informații document

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Tin-Yau Tam Multilinear Algebra (Lecture Notes)

Încărcat de

Drepturi de autor:

Formate disponibile

Multilinear Algebra1

November 30, 2011

Review of Linear Algebra

1.1 Linear extension

which are subspaces. It is known that T is injective if and only if Ker T = 0.

Theorem 1.1.1. Let E = {e1 , . . . , en } be a basis of V and let w1 , . . . , wn W .

In other words, a linear map is completely determined by the images of basis

i.e., v = T 1 w1 + 2 T 1 w2 for all w1 , w2 W and v V . We will simply write

6. Show that if T Hom (V, U ), S Hom (U, W ) are invertible, then ST is

7. Show that A Cmn is invertible only if m = n, i.e., A has to be square.

8. Prove that if A, B Cnn such that AB = In , then BA = In .

Solutions to Problems 1.1

1. Let {e1 , . . . , en } and {f1 , . . . , fm } be bases of V and W respectively. Then

2. From definition rank T dim W . From Theorem 1.1.2 rank T dim V .

3. Since Im ST Im S, rank ST rank S. By Theorem 1.1.2 rank ST =

4. Suppose that S and S 0 Hom (V, W ) are inverses of T Hom (V, W ).

5. One implication follows from Problem 4. If dim V = dim W , let E =

6. (T 1 S 1 )(ST ) = IV and (ST )(T 1 S 1 ) = IW .

7. From Problem 4, a matrix A is invertible only if A is square. The second

8. If AB = I, then rank AB = n so rank A = n(= rank B) by Problem

1.2 Matrix representations of linear maps

Let T Hom (V, W ) with bases E = {e1 , . . . , en } for V and F = {f1 , . . . , fm }

The matrix A = (aij ) Cmn is called the matrix representation of T with

Suppose that A and B are similar, i.e., B = R1 BR. Let E be a basis of V .

So functions : Cnn C that take constant values on similarity orbits

Solutions to Problems 1.2

1.3 Inner product spaces

Theorem 1.3.3. Let E = {e1 , . . . , en } be an orthonormal basis of V . For any

Denote by hv1 , . . . , vk i the span of the vectors v1 , . . . , vk .

Theorem 1.3.4. (Gram-Schmidt orthogonalization) Let V be an inner product

It is direct computation to check that {e1 , . . . , en } is the desired orthonormal

The inner product on Cn : for any x = (x1 , . . . , xn )T and y = (y1 , . . . , yn )T

kxk2 := (x, x)1/2 .

1. Let E = {e1 , . . . , en } be a basis of V and v V . Prove that v = 0 if and

2. Show that each orthogonal set of nonzero vectors is linearly independent.

4. Show that an inner product is completely determined by an orthonormal

5. Show that (A, B) = tr (B A) defines an inner product on Cmn , where

6. Prove |tr (A B)| tr (A A)tr (B B) for all A, B Cmn .

Solutions to Problems 1.3

1. v = 0 (v, u) = 0 for allPu V (v, ei ) = 0 for all i. Conversely

3. Follows from Theorem 1.3.3.

4. Follows from Problem 3.

6. Apply Cauchy-Schwarz inequality on the inner product defined in Problem

Proof. By Theorem 1.3.4, let E = {e1 , . . . , en } be an orthonormal basis of V .

Uniqueness follows from Problem 1.

Theorem 1.4.2. Let E = {e1 , . . . , en } and F = {f1 , . . . , fm } be orthonormal

Notice that if E and F are not orthonormal, then [T ]E F

(a) (T v, w) = 0 for all v V and w W if and only if T = 0.

2. Show that (ST ) = T S where T Hom (V, W ) and S L(W, U ) and

3. Let E and F be orthonormal bases of inner product space V . Prove that

4. Let G be a basis of the inner product space V . Let T End V . Prove

5. Let V, W be inner product spaces. Prove that if T Hom (V, W ), then

6. The adjoint : Hom (V, W ) Hom (W, V ) is an isomorphism.

Solutions to Problems 1.4

1. (a) (T v, w) = 0 for all v V and w W T v = 0 for all v V , i.e.,

2. Since (v, T S w)V = (T v, S w)V = (ST v, w)V , by the uniqueness of ad-

3. Notice that I = I where I := IV . So from Theorem 1.4.2 ([I]F

4. (Roy and Alex) Let E be an orthonormal basis of V . Let P = [IV ]E G

6. It is straightforward to show that : Hom (V, W ) Hom (W, V ) is a

(T w, v) = (v, T w)V = (T v, w)V = (w, T v)

Then for any S, T Hom (V, W ), S = T S = T .

1.5 Normal operators and matrices

So air = 0 for all i 6= r. Then apply it on the normal A .